CPD-FCOS: A Scale-Isolated P2 Pathway for UAV Small-Object Detection

Highlights

What are the main findings?

• A real stride-4 detection level supplies most of the accuracy gain in dense, multi-class VisDrone scenes, and also all of the added computation: $+2.98$ mAP for $2.05\times$ the end-to-end latency.
• High-resolution capacity and constrained supervision are measurably complementary: added capacity alone buys recall without precision, and only constrained positive selection converts it into average precision.

What are the implications of the main findings?

• Fine-scale detection gains should be reported by object size: the advantage here grows as objects shrink and disappears on large objects.
• A stride-4 path retains single-forward inference, but its cost falls on the detection head rather than on post-processing, which is where efficiency work should be directed.

Abstract

Small objects in UAV imagery fail for two coupled reasons. Repeated downsampling destroys their evidence—$60.5\%$ of VisDrone instances are COCO-small and the median object spans 26 pixels, so a typical target covers fewer than four cells at stride 8—and the in-box assignment rule then places most of their positive locations on object borders, where supervision is unreliable. The two reinforce each other: a finer level creates more candidate locations, but they help only if the assignment rule can select the reliable ones. We therefore design representation and supervision together in CPD-FCOS (Center-constrained P2 Detail-enhanced FCOS): a real stride-4 $P_2$ level in a bidirectional pyramid, localized residual refinement at that level, and center-constrained, scale-bounded positive selection. Controlled experiments separate the two. On VisDrone, the stride-4 level supplies $+2.98$ mAP at $2.05\times$ the end-to-end latency, while the refinement branch adds $+0.55$ mAP for $2.8\%$ more parameters. Resolved by object size, the branch alone raises small-object recall but leaves AP unchanged; only constrained assignment converts that capacity into precision. The advantage is scale-specific, growing as objects shrink and reversing on the largest ones, and it is strongest in dense multi-class scenes—on vehicle-only and medium-object benchmarks, the pathway shifts the operating point rather than dominating. The result is a single-forward, reproducible reference design with an explicitly measured accuracy–efficiency trade-off.

1. Introduction

Unmanned aerial vehicles (UAVs) offer strong maneuverability, wide observation coverage, and low deployment cost, and have become important sensing platforms for traffic monitoring, disaster search and rescue, power-line inspection, precision agriculture, and environmental monitoring. Object detection provides the basis for downstream multi-object tracking, counting, behavior analysis, and autonomous decision-making, and its reliability directly affects the practical usability of UAV vision systems. Compared with ground-view natural images, aerial images exhibit pronounced imbalance at the scale, spatial, and semantic levels [1]: flight altitude and the top-down viewpoint cause a large fraction of objects to occupy only a few to a few dozen pixels; instances such as vehicles and pedestrians often cluster non-uniformly and are accompanied by occlusion; and similar ground textures, motion blur, and a wide scale range further reduce the signal-to-noise ratio of features. These difficulties are especially prominent in benchmarks such as VisDrone [2] and UAVDT [3], causing generic detectors to suffer from missed detections, class confusion, and unstable boundary regression in UAV scenarios. Figure 1 makes the two coupled failure modes concrete on real data.

In recent years, the two-stage Faster R-CNN [4], the one-stage RetinaNet [5], and the anchor-free FCOS [6] have advanced generic object detection, yet directly applying these paradigms to UAV imagery still faces two tightly coupled difficulties. First, successive downsampling in the backbone rapidly compresses the spatial support of tiny objects; even when an FPN [7] propagates high-level semantics, the edges, contours, and local contrast of the finest level may still be overwhelmed by background responses during cross-scale fusion [8,9]. Second, dense prediction relies on per-location label assignment. Small objects have inherently limited candidate locations; a slight coordinate shift can induce a large IoU fluctuation, and supervising ambiguous near-boundary locations identically to center locations tends to aggravate the mismatch between classification and localization quality [10,11]. These problems are coupled: preserving a stride-4 feature creates more candidate locations, but those locations are useful only when the assignment rule supplies sufficiently reliable targets. Consequently, representation and supervision should be designed together rather than treated as independent add-ons.

Existing UAV-detection research alleviates these difficulties along two broad paths, reviewed in Section 2. One family locates object-dense regions and then crops, rescales, or super-resolves them, which increases the pixel support of objects at the cost of region generation, several forward passes, and cross-crop merging. The other strengthens multi-scale representation or reorganizes positive/negative assignment within a single forward pass, which suits resource-constrained deployment but usually optimizes features, label assignment, or detection quality separately. This leaves an important design question: how can high-resolution detail preservation and reliable supervision be organized as one coherent single-stage pathway, without crop generation or repeated inference? CPD-FCOS addresses this question by treating the finest feature level and its positive locations as a coupled design problem.

Motivated by this, we propose Center-constrained P2 Detail-enhanced FCOS (CPD-FCOS), a scale-isolated pathway that explicitly couples fine-detail representation with positive-location quality for tiny-object detection in UAV imagery. The model adopts a hierarchical ConvNeXt [12] backbone and a bidirectional FPN/PAN [7,13] to build $P_2$–$P_5$ multi-scale features, without relying on an extra cropper or super-resolution network. The pathway isolates the tiny-object scale through three coordinated decisions: the representation keeps a real stride-4 detection level, the $P_2$ feature is refined by a localized residual branch, and the training targets are selected by a center-constrained, scale-bounded positive-selection rule. At inference, the classification score is multiplied by the centerness prediction, providing a quality-aware confidence used for ranking and non-maximum suppression.

The main contributions of this paper are summarized as follows:

• We isolate the contribution of the stride-4 level itself. Under an identical backbone, neck design, head, schedule, and seed, replacing the $P_2$–$P_5$ pyramid with $P_3$–$P_5$ costs $2.98$ mAP and $4.90$ ${\mathrm{AP}}_S$ on VisDrone and saves $2.05\times$ end-to-end latency, so the level is expensive and decisive rather than incremental. Inference remains a single forward pass, with no crop proposal or super-resolution stage.
• We design a localized $P_2$ refinement branch with two Conv–GN–SiLU blocks and a residual shortcut. The branch adds approximately $1.2$ M parameters ($2.8\%$ of the detector) and deliberately concentrates capacity at the spatial scale, where tiny UAV targets are most vulnerable to detail dilution. Measured on the trained model, the residual it adds is comparable in magnitude to the base feature and only mildly object-selective, which explains why it needs constrained supervision to be useful.
• We configure standard FCOS primitives—in-box testing, center sampling, level ranges and minimum-area conflict resolution—for a $P_2$–$P_5$ pyramid, and quantify the resulting complementarity directly rather than inferring it from aggregate AP. Resolved by object size, the refinement branch alone raises small-object recall by $5.6$ points while leaving AP unchanged, whereas the two together also raise precision by $5.4$ points. We make no claim of novelty for centerness-based score calibration, which is unchanged from FCOS.

2. Related Work

2.1. UAV Small-Object Detection

Objects in UAV images exhibit typical properties such as small scale, non-uniform distribution, and complex background; related methods can be grouped into region-focusing and end-to-end representation-enhancement categories [1]. Region-focusing methods improve the effective resolution by removing object-free background, differing mainly in how the regions are found: from predicted object clusters [14], from a density map [15], from adaptive selection combined with local super-resolution [16], or by packing clustered foreground onto a unified canvas [17]. Such methods are effective on ultra-high-resolution images, but the pipeline gains extra stages: region generation, resampling, and result fusion.

End-to-end methods focus on improving the learning of objects at different scales within a single forward pass. SDPDet [18] models non-uniform object distributions via a scale-separated activation pyramid and dynamic proposals; a high-resolution feature pyramid for the UAV view [9] specifically enhances the shallow features where small objects reside. More recent UAV detectors broaden this accuracy–efficiency trade-off through wavelet convolution for embedded inference [19], joint treatment of small instances and long-tailed classes [20], lightweight UAV-oriented architectures [21], receptive-field interaction for dense and degraded scenes [22], and coordinated label assignment with aerial-specific augmentation [23]. Compared with region-cropping methods, our approach retains the single-stage dense-prediction pipeline over the full image and concentrates the study on two mutually dependent aspects: whether the highest-resolution feature can preserve effective detail, and whether the limited candidate locations can receive reliable supervision.

A complementary line of work attacks the same problem from the spectral side, and two of its ideas bear on the design studied here. DirectNet [24] notes that reconstruction-based anomaly detectors fit anomalies as readily as background; it therefore erases the center block from the receptive field and predicts each pixel from an outer window only, forcing the reconstruction towards the background distribution and amplifying the error at anomalous targets. The mechanism is not extra capacity but a deliberate restriction on the evidence the model may use—the same kind of decision as the center-constrained assignment of Section 3.3. VHCDN [25] transfers object priors from abundant visible-light imagery to label-scarce hyperspectral scenes, and reports that reconstructed residual representations suppress background and emphasize foreground in both domains.

That observation is a useful reference point for our own measurements: the residual branch here is only mildly object-biased (Section 4.7), which is why it needs constrained supervision to become useful, and a module designed explicitly for foreground emphasis suggests that a more selective refinement is achievable. More broadly, these methods add separability along the channel dimension whereas the present work adds it along the spatial dimension; the two are composable, since nothing in the pathway below depends on the channel semantics of the input. We do not test this here: all three benchmarks used in this study are RGB.

2.2. Feature Pyramids and High-Resolution Detail Modeling

FPN [7] constructs semantically consistent multi-scale features with a top-down path and lateral connections, and has become a basic component of modern detectors. PANet [13] adds a bottom-up path to shorten the distance over which shallow localization information propagates to high levels; NAS-FPN [26] optimizes cross-scale connections via neural architecture search; and EfficientDet [27] builds a bidirectional feature pyramid with learnable weights to improve the accuracy–computation trade-off. These methods mainly focus on generic scale fusion, whereas UAV small objects often concentrate on high-resolution levels such as $P_2/P_3$, whose effective representation is more susceptible to downsampling and background-texture interference.

Directly enlarging the input or maintaining dense high-resolution features for a long time can alleviate detail loss, but the computation and memory overhead grow rapidly with spatial size. QueryDet [8] uses low-resolution features to produce sparse queries and then computes high-resolution locations on demand; HRFPN [9] designs a high-resolution pyramid for UAV images to strengthen shallow semantics and cross-level interaction. SET [28] analyzes tiny-object evidence in the frequency domain and combines background smoothing with training-time perturbation, while Bian et al. [29] use feature information and position-dependent Gaussian supervision to reactivate weak tiny-object regions. These studies show that retaining a fine feature level is necessary but not sufficient for discriminative detail. Locate then Calibrate [30] also combines a $P_2$ detection head with adaptive scale calibration, reinforcing that the contribution must lie in how $P_2$ is organized and supervised, rather than in its presence alone. We adopt a more localized strategy: the hierarchical ConvNeXt backbone directly provides $C_2$–$C_5$ features, and only the fused $P_2$ undergoes residual local refinement, leaving the detection paths of the other scales unchanged, thereby concentrating extra representational capacity at the level most sensitive to tiny objects.

Several of these systems report stronger VisDrone numbers than the detector studied here. We list them alongside our own in Section 4.5 with the factors that make them incomparable—backbone, input size, pretraining corpus, and whether inference uses region crops or several forward passes—and we do not reproduce them in our codebase. The evidence for the design claims in this paper comes exclusively from the controlled experiments of Section 4.7, Section 4.8, Section 4.9 and Section 4.10, in which a single factor changes at a time.

2.3. Label Assignment and Detection-Quality Modeling

Label assignment determines how candidate locations in a dense detector are matched to ground truth, and is a key link connecting feature learning with the detection objective. FCOS [6] treats in-box locations as candidate positives and suppresses low-quality predictions far from object centers via center sampling and centerness. Later rules make the assignment adaptive: from IoU statistics [31], as optimal transport [32], or through task-aligned learning [33]. A parallel line encodes localization quality into the classification target [34,35], while instance-level balancing [36] and instance-wise quality distributions [37] address imbalance directly. These works make clear that instance-level balancing and quality modeling are established ideas; a new weighting mechanism must therefore specify whether it changes the assigned set, its quality distribution, or only its aggregate loss mass.

Tiny objects make label assignment more sensitive: a small pixel shift of the bounding box can cause a drastic IoU change, and each instance has very few available feature points. NWD-RKA [10] models bounding boxes as 2D Gaussians and performs ranking assignment with a normalized Wasserstein distance; RFLA [11] constructs a hierarchical assignment strategy from Gaussian receptive fields. DALA [38] further separates dense and sparse regions and adapts the number and type of matches to scene density. BCDet [39] models receptive fields and ground-truth boxes as Gaussians, then ranks candidate samples by normalized distribution similarity and joint classification–localization quality. CASA-Net [23] enforces a minimum positive count for tiny instances and relaxes their spatial tolerance. These approaches directly target insufficient or unstable positive supervision and delimit the novelty available to any new coverage-repair rule. Uncertainty-Aware Gradient Stabilization [40] analyzes the sharp localization-loss curvature of small boxes and replaces continuous regression with bounded classification-style localization and uncertainty refinement. CPD-FCOS follows a distinct route: it keeps the computationally inexpensive FCOS primitives—in-box testing, center sampling, level ranges, and minimum-area conflict resolution—but configures them as a scale-constrained positive-selection strategy for an explicit $P_2$–$P_5$ pyramid. The contribution is the resulting coupling: the added stride-4 capacity is exposed to locations selected for geometric reliability, while centerness provides a quality-aware confidence signal for ranking and suppression. This gives the detector a single, testable mechanism for connecting fine-scale representation to fine-scale supervision in UAV imagery.

3. Proposed Method

3.1. Overall Framework

The overall architecture of CPD-FCOS is shown in Figure 2 and comprises five stages: feature extraction, cross-scale fusion, detail enhancement, dense prediction, and training supervision. Given an input image, an ImageNet [41]-pretrained ConvNeXt-Tiny backbone first extracts four levels of features $\{C_2,C_3,C_4,C_5\}$ with strides $4,8,16,32$; a lightweight FPN/PAN [7,13] then aggregates top-down semantic information and bottom-up localization information to obtain $\{P_2,P_3,P_4,P_5\}$. Since tiny objects mainly depend on the highest spatial resolution level, we apply residual detail enhancement only to $P_2$ to form ${\tilde{P}}_2$, and feed $\{{\tilde{P}}_2,P_3,P_4,P_5\}$ into a shared FCOS head. Each pyramid level produces classification, bounding-box regression, and centerness predictions; the head shares the convolutional towers across scales and keeps an independent regression scale factor for each level.

Beyond the forward path, training uses center-constrained assignment: positive candidates must lie inside a ground-truth box, within a local center region, and within the scale range of the current level. The experiments call this configuration center-guided assignment (CGA); it restricts the many locations introduced by stride-4 prediction to those with stronger central evidence, and is training-only, so inference adds no crop generation, super-resolution, or additional forward pass.

Table 1. Structural distinction between standard FCOS and CPD-FCOS.

Aspect	Standard FCOS	CPD-FCOS
Feature levels	Typically $P_3$–$P_7$, strides 8–128	Real $P_2$–$P_5$, strides 4–32
Cross-scale fusion	Top-down FPN	Bidirectional FPN/PAN
Assignment	In-box candidates with center sampling and level ranges	Same primitives, configured for $P_2$–$P_5$ with compact overlapping ranges
Fine-level branch	No $P_2$-specific refinement	Two-block residual refinement only at $P_2$
Score calibration	Classification score modulated by centerness	Unchanged from FCOS
Inference flow	Single forward pass	Single forward pass

summarizes the resulting structural difference from standard FCOS.

3.2. Multi-Scale Feature Fusion

To accommodate the large-scale variation in objects in UAV images, we build a feature pyramid from the four levels output by the hierarchical backbone. For each backbone feature $C_i$, a $1\times 1$ convolution is first applied for channel alignment to obtain the lateral feature $L_i$. In top-down fusion, the high-level feature is upsampled to the spatial resolution of the adjacent lower level and, after concatenation with the lower-level lateral feature, fused by a convolution–normalization–activation unit. Let $T_i$ denote the intermediate feature of the top-down path; then

$$\begin{array}{cc}\hfill T_5& =L_5,\hfill \\ \hfill T_i& ={\eta }_i\left([L_i,\mathrm{Up}\left(T_{i+1}\right)]\right),i=2,3,4,\hfill \end{array}$$

(1)

where $[·]$ denotes channel concatenation and ${\eta }_i$ denotes two successive $3\times 3$ Conv–GroupNorm–SiLU units: the first maps the concatenated 512 channels back to 256, and the second refines the fused feature. In the bottom-up path, the finer-scale output is downsampled by a stride-2 convolution, added to the adjacent $T_i$, and refined by ${\xi }_i$:

$$P_2=T_2,P_i={\xi }_i\left(T_i+\mathrm{Down}\left(P_{i-1}\right)\right),i=3,4,5.$$

(2)

This bidirectional path lets $P_2$ obtain high-level semantics while continuing to propagate shallow localization cues to $P_3$–$P_5$. Compared with a backbone that outputs a single-stride feature, the hierarchical ConvNeXt provides a genuine high-resolution $C_2/P_2$ representation for UAV small objects; all outputs are unified to 256 channels for subsequent processing by the shared head.

3.3. Center-Constrained Assignment for the $P_2$–$P_5$ Pyramid

In dense UAV images, edge locations of small objects often lack complete object evidence and are highly sensitive to coordinate perturbations of a few pixels. If positive samples are selected only by “being inside the ground-truth box,” many boundary and overlapping-region locations may produce unstable classification and regression supervision, as shown in Figure 3a. CPD-FCOS therefore jointly applies the in-box constraint, the center neighborhood, and the level regression range, following the center-sampling recipe of FCOS [6]. Let the feature stride of the l-th level be $s_l$, the prediction location be $p=(x,y)$, and the ground-truth box be $b=(x_1,y_1,x_2,y_2)$ with center $c=(c_x,c_y)$; denote the maximum regression distance from this location to the four box edges as $d_{\max }(p,b)=\max (x-x_1,y-y_1,x_2-x,y_2-y)$. A location p becomes a candidate positive of object b only if it simultaneously satisfies

$$\begin{array}{cc}\hfill & x_1<x<x_2,y_1<y<y_2,\hfill \\ \hfill & |x-c_x|\le r{s}_l,|y-c_y|\le r{s}_l,\hfill \\ \hfill & m_l^{-}\le d_{\max }(p,b)\le m_l^{+},\hfill \end{array}$$

(3)

where r is the center-neighborhood radius coefficient, set to $1.5$ in the default configuration unless otherwise stated, and $[m_l^{-},m_l^{+}]$ is the regression range of the l-th pyramid level. If a location satisfies the conditions of multiple ground-truth objects simultaneously, it is matched to the object of smallest area; the remaining locations that do not satisfy the conditions are treated as negatives. For $P_2$–$P_5$, the ranges are $[0,32]$, $[16,64]$, $[32,128]$, and $[64,+\infty )$ pixels. Their overlap prevents a hard scale transition, while the center window limits the number of positives introduced by the dense stride-4 grid. Figure 3b illustrates the resulting candidate set. Section 4.7 measures what this rule actually does to the positive set: it removes $55.8\%$ of all positive locations while leaving the fraction of small objects with no positive at all essentially unchanged ($1.17\%\to 1.27\%$), so it redistributes supervision rather than withholding it.

3.4. Scale-Isolated $P_2$ Detail Enhancement

After bidirectional fusion, the stride-4 $P_2$ has the highest spatial resolution among all pyramid levels and is therefore crucial for objects covering only a few pixels. However, cross-scale fusion may also dilute weak contours and local contrast with background responses. To this end, we add the lightweight residual detail-enhancement branch shown at the stride-4 row of Figure 2 to the $P_2$ output. Specifically, let ${\varphi }_k(·)=\mathrm{SiLU}(\mathrm{GN}({\mathrm{Conv}}_{3\times 3}^{\left(k\right)}(·)))$; then the enhanced feature is

$${\tilde{P}}_2=P_2+{\varphi }_2({\varphi }_1(P_2)).$$

(4)

The two convolutional units keep the feature channel count and spatial size unchanged, and the identity shortcut preserves the original pyramid information and improves optimization stability. The enhancement acts only on $P_2$, leaving the $P_3$–$P_5$ paths unchanged, so the extra representational capacity is concentrated at the scale that most needs fine-grained cues, without repeatedly introducing computation on all high-level features. This branch adds about $1.2$ M parameters, increasing the model size from about $42.3$ M to $43.5$ M, a structural increment of about $2.8\%$.

For completeness, the shared FCOS head predicts raw distances ${\mathbf{z}}_l$ at level l. The implementation converts them to non-negative pixel distances as ${\widehat{\mathbf{d}}}_l$ = $s_l\mathrm{softplus}\left(q_l{\mathbf{z}}_l\right)$, where $q_l$ is a learned scalar specific to the level and $s_l$ is its stride. This parameterization preserves a shared head while allowing scale-specific regression calibration. During inference, the classification probability is multiplied by the predicted centerness before thresholding and NMS, so detections whose classification evidence is not supported by localization quality are ranked lower.

3.5. Training Objective

The classification branch uses focal loss [5], the bounding-box branch uses an IoU loss, and the centerness branch uses binary cross-entropy. Let $\mathcal{P}$ and $\mathcal{N}$ denote the sets of positive and negative locations and $N=\max \left(\right|\mathcal{P}|,1)$ the number of positives; the total training objective is

$$\mathcal{L}={\displaystyle \frac{1}{N}}\left[\sum _{j\in \mathcal{P}\cup \mathcal{N}}{\mathcal{L}}_{cls}^{\left(j\right)}+\sum _{j\in \mathcal{P}}\left({\lambda }_{reg}{\mathcal{L}}_{reg}^{\left(j\right)}+{\lambda }_{ctr}{\mathcal{L}}_{ctr}^{\left(j\right)}\right)\right],$$

(5)

where ${\lambda }_{reg}$ and ${\lambda }_{ctr}$ are both set to 1, and every positive sample contributes equally to the loss in the final reported model. Section 4.12 separately examines an area-weighted variant to test whether changing the loss weights provides additional benefit beyond the scale-isolated pathway.

4. Experiments

4.1. Datasets and Splits

We conduct experiments on three public UAV datasets. VisDrone [2] contains ten categories of urban traffic participants, characterized by dense instances, frequent occlusion, and complex backgrounds; we use the official training and validation sets for fitting and model selection, and the annotated test-dev set for the locked evaluation. UAVDT [3] focuses on vehicle detection with three categories—car, truck, and bus. Since its official split does not provide a separate validation set, we hold out six sequences from the official training sequences at the sequence level—M0210, M0902, M1002, M1006, M1304, and M1305—so as to avoid leakage of adjacent video frames across sets. UAVVaste [42] targets low-altitude UAV litter detection as a single-category task and uses the released split. We select checkpoints and hyperparameters using validation data only and keep the annotated test sets locked for the final evaluation; the train, validation, and test filename sets have zero pairwise intersections in all three datasets.

Table 2. Statistics of the train, validation, and locked test splits.

Dataset	#Classes	Split	#Images	#Boxes
		train	6471	343,197
VisDrone	10	val	548	38,759
		test	1610	75,101
		train	6471	343,197
VisDrone	10	train	19,658	342,920
UAVDT	3	val	4485	79,991
		test	16,592	375,884
		train	540	2595
UAVVaste	1	val	116	616
		test	116	507

summarizes the number of images and valid annotated boxes actually used after the same annotation filtering; the three datasets cover multi-category dense detection, extremely small vehicle detection, and single-category low-altitude detection, respectively.

4.2. Implementation Details

The backbone is an ImageNet-pretrained ConvNeXt-Tiny; $C_2$–$C_5$ are unified to 256 channels by the FPN/PAN, and both the FCOS classification and regression towers contain four $3\times 3$ Conv–GN–SiLU units. The released paper configuration fails explicitly if the requested pretrained weights cannot be loaded, preventing an unnoticed fallback to random initialization. The regression ranges of the four pyramid levels are set to $[0,32]$, $[16,64]$, $[32,128]$, and $[64,+\infty )$, and the default center-neighborhood radius coefficient is $r=1.5$. The focal-loss parameters are $\alpha =0.25$ and ${\gamma }_f=2$, and every positive sample contributes equally to the loss (no scale-aware reweighting; see the ablation). At inference, the confidence threshold and NMS IoU threshold are $0.05$ and $0.6$, and at most 1000 candidates are kept per image.

Configurations are selected under four fixed rules, applied in order. (i) Selection uses validation data only; the locked test splits are never consulted. (ii) A single default setting is shared by all three datasets, so that dataset-specific tuning cannot be mistaken for a property of the method. (iii) When a difference falls inside the multi-seed dispersion, the original FCOS default is retained rather than the nominally better value. (iv) When mAP and ${\mathrm{AP}}_S$ disagree, ${\mathrm{AP}}_S$ decides, because the study targets small objects, and the sacrificed mAP is reported explicitly. Section 4.10 applies rule (iii) to the center radius.

FCOS is used as the research platform rather than as the strongest available detector. Studying whether representation and supervision must be designed jointly requires a platform whose assignment rule can be switched on and off cleanly. In FCOS, that rule is three analytic conditions—an in-box test, a center radius and a level range—each ablatable without disturbing the others, and the level–range mechanism extends naturally to a stride-4 level. Modern YOLO detectors couple task-aligned assignment to mosaic augmentation, exponential moving averages, distribution focal loss, and a tuned loss schedule, so a factorial ablation there cannot separate the assignment rule from the training recipe. YOLO is therefore evaluated as a comparison system in Section 4.5, not used as the substrate for the ablations.

The default input size is $1024\times 1024$; training uses bf16 mixed precision and the AdamW optimizer with weight decay $0.05$, a cosine schedule decaying the learning rate to ${10}^{-6}$, and gradient-norm clipping of one. The controlled ablations use an initial learning rate of ${10}^{-4}$, a training budget of 40k steps on VisDrone and UAVDT and 12k steps on UAVVaste, a per-GPU batch size of three on VisDrone and four on the other datasets. Each reported CPD-FCOS seed is trained on one GPU. Basic augmentation includes random horizontal flipping, scale perturbation, copy-paste, and photometric perturbation. To adapt to dataset differences, the main results use the configuration selected on the validation set: UAVDT lowers the initial learning rate to $5\times {10}^{-5}$ to mitigate rapid overfitting caused by redundant video frames; UAVVaste raises the copy-paste probability to $0.8$ and the maximum number of pasted instances per image to 24, and adopts a wider scale-perturbation range of $[0.4,2.0]$ and a 15k-step budget. Unless otherwise stated, we report the checkpoint with the highest validation mAP within each run. The primary aggregate result is the mean and standard deviation across three seeds; individual test runs are retained in the machine-readable evaluation record but are not ranked in the manuscript. Dataset-specific learning rates and augmentation are held fixed across the three seeds of a dataset. Controlled ablations instead use the common recipe described above to isolate incremental component effects. The implementation uses PyTorch 2.4.0, Torchvision 0.19.0, PyTorch Lightning 2.5.2, TorchMetrics 1.8.0, and timm 1.0.27 with CUDA 11.8; experiments are executed on NVIDIA GeForce RTX 4090 D GPUs.

4.3. Evaluation Protocol

We adopt COCO [43]-style evaluation metrics, including mAP ($\mathrm{AP}@[0.5:0.95]$), ${\mathrm{AP}}_{50}$, and the area-partitioned ${\mathrm{AP}}_S/{\mathrm{AP}}_M/{\mathrm{AP}}_L$. Different detectors may use their own input-scaling strategies, but during evaluation all methods infer on the same complete split and map predictions back to the original image resolution before computing metrics against the original annotations. This protocol ensures that the COCO area partition is determined by the original object area, avoiding incomparable bias on ${\mathrm{AP}}_S$ caused by square- or aspect-preserving resizing. The detector may retain up to 1000 pre-evaluation candidates per image, whereas the reported standard COCO mAP uses the conventional maximum of 100 detections per image. The primary aggregate results use the locked test splits once after validation-based model selection. One further locked-test evaluation was carried out during revision, for the alternative center radius of Section 4.10. That configuration was chosen entirely from validation evidence; the test splits played no part in selecting it, and it did not become the reported default; we record the second evaluation here so that the count is not overstated. Cross-method comparisons and controlled ablations use the validation splits under the same coordinate protocol and are labeled separately; their data are not pooled with the locked-test results. For VisDrone, category-0 ignored regions are excluded during COCO conversion, and the evaluator does not subsequently suppress predictions that overlap those regions. The resulting metric is therefore a documented COCO-style benchmark, not the official VisDrone challenge evaluator; all locally trained methods use the same conversion, while author-reported challenge-style values are presented only as protocol-labeled context. A post hoc validation audit quantifies this protocol difference using a Python port of the public VisDrone DET toolkit, including ignored-region suppression and its 500-detection limit. On identical saved predictions, the three CPD-FCOS seeds change from $0.3264\pm 0.0008$ COCO-style mAP to $0.3383\pm 0.0010$ toolkit-style AP and the YOLOv11s reference from $0.2877$ to $0.2972$, so the margin remains similar ($3.88$ versus $4.11$ points) rather than reversing. This is a diagnostic check, not a second locked test: the checkpoints were selected on validation, and the public toolkit does not report the area-partitioned ${\mathrm{AP}}_S$ used in our primary analysis.

4.4. Locked-Test Results

Table 3. CPD-FCOS results on the locked test splits at original resolution. Each entry is the mean ± population standard deviation over seeds 42, 123, and 456; no test run is selected as “best”.

Dataset	mAP	${\mathbf{AP}}_{\mathbf{50}}$	${\mathbf{AP}}_{\mathit{S}}$	${\mathbf{AP}}_{\mathit{M}}$	${\mathbf{AP}}_{\mathit{L}}$
VisDrone	0.244 ± 0.001	0.426 ± 0.001	0.151 ± 0.001	0.342 ± 0.001	0.426 ± 0.018
UAVDT	0.163 ± 0.008	0.282 ± 0.019	0.117 ± 0.010	0.258 ± 0.010	0.260 ± 0.028
UAVVaste	0.528 ± 0.005	0.865 ± 0.012	0.124 ± 0.024	0.500 ± 0.002	0.640 ± 0.010

reports the one-time locked-test evaluation of CPD-FCOS. All three pre-specified seeds are included, with population standard deviation ($\mathrm{ddof}=0$), so the table reflects the behavior of the fixed method rather than a favorable test-run selection. Two patterns stand out. VisDrone is both strong and stable across the reported metrics, whereas UAVDT is more seed-sensitive and shows a marked validation-to-test drop. UAVVaste is stable in overall mAP, but its ${\mathrm{AP}}_S$ varies more because the test split contains only 32 small annotations. The cross-dataset pattern is therefore informative: the pathway is most reliable in dense multi-class scenes, while its transfer to vehicle-only or small-sample settings requires metric-specific interpretation.

For a test reference that does not use test data for method choice, we pre-selected the highest-validation-mAP baseline within the local comparison set: YOLOv11s for VisDrone, FCOS for UAVDT, and YOLOv8s for UAVVaste.

Table 4. Locked-test comparison with the baseline selected by validation mAP for each dataset. Both sides are now three-seed means ± population standard deviations over seeds 42, 123 and 456. Selection was completed without test metrics.

Dataset	Method	mAP	${\mathbf{AP}}_{\mathbf{50}}$	${\mathbf{AP}}_{\mathit{S}}$	${\mathbf{AP}}_{\mathit{M}}$	${\mathbf{AP}}_{\mathit{L}}$
VisDrone	YOLOv11s	0.223 ± 0.001	0.380 ± 0.001	0.120 ± 0.002	0.336 ± 0.000	0.471 ± 0.008
	CPD-FCOS	0.244	0.426	0.151	0.342	0.426
UAVDT	FCOS	0.167 ± 0.002	0.289 ± 0.003	0.129 ± 0.003	0.270 ± 0.004	0.126 ± 0.015
	CPD-FCOS	0.163	0.282	0.117	0.258	0.260
UAVVaste	YOLOv8s	0.517 ± 0.005	0.817 ± 0.004	0.059 ± 0.009	0.472 ± 0.008	0.648 ± 0.003
	CPD-FCOS	0.528	0.865	0.124	0.500	0.640

The baselines were retrained for this revision with the same three seeds as CPD-FCOS. Their training scripts had no seed control, so seeding was added; for FCOS, the small per-image batch also made SGD diverge at the original learning rate on all three seeds, and gradient clipping—which the proposed model already uses—was added to stabilize it.

places the three-seed test results of these validation-selected baselines alongside the corresponding three-seed CPD-FCOS results. The VisDrone result is the clearest signal: CPD-FCOS gains $2.1$ mAP points and $3.1$ ${\mathrm{AP}}_S$ points against a baseline dispersion of $0.001$ and $0.002$, showing that the high-resolution pathway survives the locked split rather than merely fitting the validation distribution. UAVDT tells a different story. With the baseline repeated, FCOS is now ahead on every metric except ${\mathrm{AP}}_L$: $0.167$ versus $0.163$ mAP and $0.129$ versus $0.117$ ${\mathrm{AP}}_S$. The single FCOS run reported in the previous version of this table ($0.161$ mAP) was at the low end of its own seed distribution, so repeating it removed a small apparent advantage of the proposed method rather than creating one. In contrast, UAVVaste shows a $1.1$-point mAP gain over YOLOv8s, although the apparent $6.5$-point ${\mathrm{AP}}_S$ gain should be read alongside the small number of small annotations.

4.5. Validation Comparison with Mainstream Detectors

To validate the effectiveness of CPD-FCOS, we compare it with eight representative detectors: Faster R-CNN [4], RetinaNet [5], FCOS [6], the Faster R-CNN v2 and RetinaNet v2 model variants from Torchvision, a Faster R-CNN with a MobileNetV3 [44] backbone, and the real-time detectors YOLOv8s [45] and YOLOv11s [46]. The Torchvision detectors all use ImageNet-pretrained backbones and are trained for 60k steps on VisDrone and UAVDT and 15k steps on the smaller UAVVaste. The YOLO models use their official small configuration and 60-epoch training pipeline. All baselines are trained and evaluated at the same input size as CPD-FCOS: the Torchvision detectors use min_size = 1024, max_size = 1707, and the YOLO models use imgsz = 1024 at both training and inference rather than the Ultralytics default of 640; the per-method sizes are listed in

Table 5. Validation-set comparison with mainstream detectors under the shared evaluation protocol (original resolution, mAP/${\mathrm{AP}}_S$). CPD-FCOS reports the three-seed mean; baseline entries are single runs. Every method is trained and evaluated at the same input size. The strongest baseline of each dataset was retrained with the same three seeds as CPD-FCOS and is reported as mean ± population standard deviation; the remaining entries are single runs. The last row is not part of the comparison: it retrains YOLOv8s under the full Ultralytics default to quantify what the shared 1024 setting is worth to a baseline.

Method	Input	VisDrone		UAVDT		UAVVaste
		mAP	${\mathbf{AP}}_{\mathit{S}}$	mAP	${\mathbf{AP}}_{\mathit{S}}$	mAP	${\mathbf{AP}}_{\mathit{S}}$
Faster R-CNN	1024/1707	0.256	0.181	0.215	0.188	0.469	0.145
RetinaNet	1024/1707	0.191	0.106	0.224	0.273	0.454	0.103
FCOS	1024/1707	0.231	0.148	$0.224\pm 0.002$	$\mathbf{0.282}\pm \mathbf{0.030}$	0.456	0.130
Faster R-CNN v2	1024/1707	0.241	0.164	0.224	0.190	0.451	0.124
RetinaNet v2	1024/1707	0.194	0.105	0.214	0.197	0.450	0.106
Faster R-CNN (MobileNet)	1024/1707	0.149	0.064	0.180	0.160	0.336	0.017
YOLOv8s	${1024}^2$	0.286	0.194	0.225	0.194	$0.482\pm 0.003$	$0.166\pm 0.014$
YOLOv11s	${1024}^2$	$0.288\pm 0.004$	$0.193\pm 0.004$	0.195	0.173	0.477	0.209
CPD-FCOS (mean)	${1024}^2$	0.326	0.251	0.261	0.287	0.488	0.196
YOLOv8s (Ultralytics default)	${640}^2$	0.211	0.112	–	–	–	–

and the configuration files are released with the code. Retraining YOLOv8s under the full Ultralytics default (640 throughout) lowers its VisDrone validation mAP from $0.286$ to $0.211$ and its ${\mathrm{AP}}_S$ from $0.194$ to $0.112$, so the shared 1024 setting is favorable to the baselines rather than to the proposed method. All methods use the same complete validation split and follow the original-resolution evaluation protocol of the previous subsection. Since each detector keeps its own optimizer, training recipe, and input preprocessing, this experiment measures practical system-level performance, while the incremental component evidence is given by the subsequent full-factorial ablations and sensitivity studies. A post hoc loader audit found that the historical UAVDT Torchvision runs omitted 314 training frames that contained no valid boxes, whereas CPD-FCOS retained those negative-only frames. The released loader now preserves them, but the checkpoint reported here predates that fix. We therefore treat the UAVDT Torchvision entries as non-controlled references and do not attribute their difference from CPD-FCOS solely to architecture.

Three kinds of evidence appear in this section and are kept separate throughout. Controlled comparisons share one codebase, optimizer, schedule, data pipeline and evaluation protocol, and vary a single factor; Section 4.7, Section 4.8, Section 4.9 and Section 4.10 are of this kind and are the sole basis for the design claims. System-level references are independent implementations kept at their own official recipes with only the input size and evaluation protocol unified; Table 5 is of this kind and locates the method in a performance range without supporting component-level conclusions. Reported values are taken from other papers and used only as protocol-annotated context, never for ranking.

Table 5 uses the three-seed CPD-FCOS mean as the primary value; the competing detectors are single-run references, and CPD-FCOS stability is reported separately in Section 4.12. The results again separate the datasets rather than producing one uniform ranking. On VisDrone, CPD-FCOS reaches $0.326$ mAP, 3.8 points above YOLOv11s, and obtains the highest ${\mathrm{AP}}_S$ in the table. This is precisely the regime targeted by the pathway: dense multi-class scenes combine tiny instances, class ambiguity, and background clutter, making both fine detail and reliable positive locations valuable. On UAVDT, the margin is narrower: CPD-FCOS leads the reproducible FCOS reference by $3.7$ mAP points, while the two ${\mathrm{AP}}_S$ values ($0.287$ against $0.282\pm 0.030$) differ by less than the seed dispersion of the baseline, so the small-object comparison on this dataset is not resolved by the available evidence. The FCOS entry deserves a comment. The single run reported previously reached $0.263$ mAP, higher than any of fifteen seeded repetitions across three learning rates, two batch sizes and gradient clipping on or off; the best reproducible configuration averages $0.224\pm 0.002$. We report the reproducible mean and record the earlier value so the discrepancy is visible rather than silently corrected. The same checkpoint sits at the low end of its seed distribution on the locked test split ($0.161$ against $0.167$), so it was favorable on validation and unfavorable on test—the instability that repeating a baseline is meant to expose. UAVVaste offers a complementary case in which CPD-FCOS remains competitive overall, but the small-object subset is too limited to support a broad small-object claim.

The generic-detector comparison places CPD-FCOS in a competitive range, while recent UAV-specific systems provide an important external reference.

Table 6. Author-reported VisDrone2019 validation results from recent UAV-specific detectors. “NR” means that the item was not reported clearly enough for a controlled comparison. CPD-FCOS is included only as a protocol reference; cross-paper values were not reproduced in our codebase. AD-Det is reported at two backbones because its paper reports both; the ResNet-50 row is the configuration whose ${\mathrm{AP}}_S$ is available. Its inference column counts one coarse pass plus $N=4$ subregion passes.

Method	Year	Input	Pretraining	Inference	AP	${\mathbf{AP}}_{\mathit{S}}$
AUHF-DETR-M [19]	2025	${640}^2$	NR	single pass	0.309	0.261
SARC-DETR [22]	2026	${1024}^2$	COCO	single pass	0.348	0.248
AD-Det (ResNet-50) [20]	2025	$1333\times 800$	ImageNet	1 + 4 passes	0.353	0.280
AD-Det (ResNeXt-101) [20]	2025	$1333\times 800$	ImageNet	1 + 4 passes	0.370	NR
CPD-FCOS (3-seed mean)	2026	${1024}^2$	ImageNet	single pass	0.326	0.251

therefore reports protocol-labeled values from the original papers. AUHF-DETR-M uses a smaller input and emphasizes embedded efficiency, whereas SARC-DETR uses a 1024-pixel setting and a COCO-pretrained transformer. SARC-DETR reports higher overall AP than CPD-FCOS, and AUHF-DETR-M reports a higher ${\mathrm{AP}}_S$. Region-focusing detectors report stronger VisDrone accuracy than the single-forward configuration studied here, and we state the gap numerically rather than qualitatively. AD-Det [20] reports $0.353$ AP with a ResNet-50 backbone and $0.370$ with ResNeXt-101 on the same validation split, against $0.326$ here, and $0.280$ ${\mathrm{AP}}_S$ against $0.251$; its own two-stage baseline, which already crops each image into four uniform parts, reports $0.331$ AP and $0.243$ ${\mathrm{AP}}_S$. The margin is therefore real, and we acknowledge it plainly. It is obtained through a largely complementary mechanism: object-dense regions are located, cropped and re-processed at higher effective resolution, which costs one coarse pass plus four subregion passes per image against the single pass measured in Section 4.13. The contribution here lies elsewhere—in isolating, under controlled conditions, how a stride-4 level and its supervision interact inside a single forward pass, and in placing that interaction against eight established detectors under one shared evaluation protocol. The two directions are compatible rather than competing: a region-focusing stage could be added on top of this pathway. Because training data handling, pretraining, and post-processing are not controlled across publications, the cross-paper values are not used for a formal ranking. The primary evidence for CPD-FCOS remains the local factorial comparison and the cross-dataset diagnosis under a documented implementation.

A cross-paper number cannot be audited, so we additionally reimplemented the AD-Det pipeline in this codebase and evaluated it on the same validation split with the evaluator used everywhere else in this paper.

Table 7. AD-Det reimplemented in this codebase and evaluated on the VisDrone validation split under the protocol of this paper (COCO AP, at most 100 detections per image). “Reported” values are from the original paper, whose own baseline already crops each image into four uniform parts. The reimplementation follows the published recipe (GFL, ResNet-50, $1333\times 800$, 12 epochs, SGD $0.01$, decay at epochs 8 and 11, $N=4$ subregions, fusion NMS $0.5$); the DCC module was not reimplemented. The reimplementation is below the published figures and is reported as such; it does not license any ranking claim against AD-Det.

Configuration	Passes/Image	Reported AP	Reproduced AP
GFL, no second stage	1	NR	0.210
GFL + four uniform crops (their baseline)	1 + 4	0.331	0.293
+ASOE, shared-weight fine detector	1 + 4	NR	0.217
+ASOE, dedicated fine detector	1 + 4	0.353	0.278
+ASOE + DCC (full AD-Det)	1 + 4	0.359	not reimplemented
CPD-FCOS (this paper), single pass	1	–	0.326

reports the outcome, including the parts that did not reproduce. The region-focusing mechanism itself replicates clearly: adding a cropped second stage with a dedicated fine detector raises the same base detector from $0.210$ to $0.293$ AP, an $8.3$-point gain that no single-pass change in this paper approaches. Two things did not replicate. The absolute level is $7.5$ points below the published figure ($0.278$ against $0.353$), which we attribute to the base detector rather than to the second stage, since our plain GFL is correspondingly weak. And the adaptive subregions did not beat the four uniform crops that AD-Det uses for its own baseline ($0.278$ against $0.293$ in our hands, where the published ordering is the reverse); in our reimplementation the clustered regions leave $7.4\%$ of object centers outside every subregion, whereas quadrants tile the image exactly. We therefore treat Table 6, which uses the authors’ own numbers, as the comparison of record, and this table as evidence about mechanism and cost only. In particular, the CPD-FCOS row below must not be read as outperforming AD-Det; it is placed there only as the single-pass reference point against which the cost of the extra passes should be judged. That cost is substantial: on one GPU, the same reimplemented base detector takes $0.033$ s per image with a single pass and $0.277$ s once the four subregion passes are added.

Figure 4 shows the corresponding precision–recall curves; each curve is the seed-42 run of the configuration reported in Table 5. On VisDrone, the CPD-FCOS curve is above the reference from recall $0.2$ upward, with the gap widening to more than 20 precision points near recall $0.7$, and it extends to a higher maximum recall ($0.94$ against $0.91$); below recall $0.2$, the two lie within about one precision point of each other. On UAVDT, the CPD-FCOS curve is above the reproducible FCOS reference at every recall, by 8 to 15 precision points below recall $0.55$, while above recall $0.6$, the two nearly coincide, and FCOS reaches a marginally higher maximum recall ($0.88$ against $0.87$). On UAVVaste, the two are close throughout.

Figure 5, Figure 6 and Figure 7 complement Table 5 with qualitative comparisons on the three datasets. Blue boxes denote ground truth, green boxes denote CPD-FCOS predictions, and orange/red boxes denote competing-detector predictions. The displayed dataset indices were fixed before visual inspection and are recorded together with the image paths, checkpoint paths, and checkpoint SHA-256 hashes in Supplementary Files S2–S4. These examples do not replace the quantitative comparison on the complete validation split.

4.6. Remote-Sensing Scenario Analysis

The three benchmarks represent different remote-sensing operating regimes, and their differences help explain the quantitative pattern. VisDrone is a dense urban monitoring benchmark with ten categories and frequent co-occurrence of pedestrians, vehicles, and non-motor objects. This is where the scale-isolated pathway is most effective: CPD-FCOS leads both mAP and ${\mathrm{AP}}_S$ in Table 5, and the qualitative examples in Figure 5 show better retention of adjacent tiny instances. In other words, the gain is not simply a consequence of adding more pyramid levels; it appears when fine detail and crowded positive-location competition occur together.

UAVDT represents a narrower vehicle-monitoring transfer setting. Its validation objects are dominated by small vehicles, but the classes are fewer, and the strong FCOS reference remains better on ${\mathrm{AP}}_S$. This contrast is useful operationally: traffic-monitoring deployments that prioritize extremely small vehicle recall should track ${\mathrm{AP}}_S$ and recall in addition to mAP. The controlled factorial study in Section 4.8 reinforces this reading: on UAVDT, the full CGA + $P_2$ pathway gives the strongest small-object AP among our variants ($0.3273$), while a $P_2$-only variant gives higher mAP. UAVVaste instead resembles low-altitude environmental monitoring, where many targets are medium or large trash objects and the small subset is limited. Here the method is competitive in mAP, but the interpretation of ${\mathrm{AP}}_S$ is less stable because few validation targets fall into that bin.

Aggregate mAP is too coarse to delimit where the method applies, so

Table 8. Per-class AP on the VisDrone validation split at original resolution, against the validation-selected reference detector.

Method	Pedes.	People	Bicycle	Car	Van	Truck	Tricy.	Awn.-Tri.	Bus	Motor
YOLOv11s	0.275	0.171	0.136	0.603	0.331	0.312	0.205	0.120	0.471	0.272
CPD-FCOS	0.322	0.221	0.170	0.631	0.408	0.340	0.229	0.121	0.516	0.296

and 

Table 9. VisDrone validation performance resolved by object size ($\sqrt{\mathrm{area}}$ in pixels). Recall is measured at IoU $=0.5$.

$\sqrt{\mathbf{Area}}$	Recall (CPD-FCOS)	Recall (YOLOv11s)	AP (CPD-FCOS)	AP (YOLOv11s)
<16	0.740	0.367	0.157	0.097
16–32	0.838	0.660	0.313	0.255
32–64	0.883	0.801	0.408	0.406
≥64	0.921	0.851	0.503	0.514

resolve the VisDrone comparison by class and by object size. CPD-FCOS leads the validation-selected YOLOv11s reference on all ten classes, and the margin tracks how small and how crowded a class typically is: it is largest for van ($+7.8$), people ($+5.0$), pedestrian ($+4.7$) and bus ($+4.4$), and smallest for awning-tricycle ($+0.2$), where the two detectors are effectively tied, and for tricycle and motor (both $+2.4$).

The size-resolved view is more informative still. Recall at IoU $=0.5$ improves by $37.3$ points on objects below 16 pixels, by $17.7$ points between 16 and 32, by $8.2$ points between 32 and 64, and by $7.1$ points above 64; average precision follows the same ordering and reverses on the largest bin, where the reference is $1.1$ points better. The advantage is therefore scale-specific rather than a general capacity effect, which is what a scale-isolated pathway should produce, and it also marks the regime where the method offers nothing.

UAVDT separates the two detectors differently. Measured against the reproducible FCOS reference, CPD-FCOS has slightly lower recall on the two smallest bins ($0.694$ against $0.700$ below 16 pixels and $0.728$ against $0.743$ between 16 and 32) and higher recall on the two largest ($0.603$ against $0.539$ and $0.614$ against $0.543$), while its AP and its precision at IoU $=0.5$ are higher in all four bins; the margin is by far the largest in the smallest bin ($0.386$ against $0.170$ AP and $0.609$ against $0.344$ precision). Per class, it leads on truck ($0.201$ against $0.106$) and bus ($0.065$ against $0.049$) and trails only on car ($0.501$ against $0.515$), the class that dominates the dataset. The pattern on this dataset is therefore one of precision rather than recall: the pathway converts the stride-4 capacity into cleaner detections at a given recall, and the residual FCOS advantage is confined to recall on the smallest instances.

4.7. Mechanism Analysis

Aggregate AP shows that the two components help, not why. This subsection replaces that inference with three direct measurements, all taken on the VisDrone validation split under the protocol of the previous subsections.

What the assignment rule does to supervision. Figure 8 audits the positive set over 256 validation images. Center-constrained selection removes more than half of all positive locations ($\mathrm{525,975}\to \mathrm{232,373}$) and halves the median count per small object ($17\to 9$), yet the fraction of small objects left with no positive at all barely moves ($1.17\%\to 1.27\%$). The rule therefore redistributes supervision rather than withholding it: the share of positives lying in the outer band of their object, $\delta =\max \left(\right|x-c_x|,|y-c_y\left|\right)/(\mathrm{short}\mathrm{side}/2)>0.6$, drops from $77.8\%$ to $57.8\%$ for small objects and from $64.1\%$ to $27.7\%$ for large ones. The residual $57.8\%$ is itself informative: the window $r{s}_l$ is an absolute distance, so it constrains the smallest objects proportionally less than the largest ones, which is also why the detector is sensitive to r (Section 4.10).

What the $P_2$ branch adds. Figure 9 shows $\parallel P_2{\parallel }_2$, $\parallel {\tilde{P}}_2{\parallel }_2$ and the residual magnitude produced by the trained branch. The residual is not a small perturbation: its magnitude reaches $0.92$–$0.95$ of the base feature. It is biased towards objects but far from exclusive to them, with mean magnitude inside ground-truth boxes only $1.23$ times the background level. The branch supplies capacity broadly; it does not by itself decide where that capacity is useful.

Why the two are complementary. Figure 10 resolves the factorial variants by object size. On the smallest bin, the $P_2$ branch alone raises recall from $0.626$ to $0.682$ but leaves AP unchanged at $0.132$, with precision essentially flat—the additional detections are matched by additional false positives. Constrained assignment alone lifts precision to $0.338$ and AP to $0.146$. Only the combination reaches $0.737$ recall, $0.349$ precision and $0.156$ AP. This is the mechanism behind the factorial result in

Table 10. Full-factorial ablation of center-constrained assignment (CGA) and $P_2$ refinement under the fixed uniform-loss recipe. A check mark indicates that the corresponding component is enabled, whereas an en dash indicates that it is disabled. Metrics are evaluated at the original image resolution.

Dataset	CGA	${\mathit{P}}_{\mathbf{2}}$ Refine	mAP	${\mathbf{AP}}_{\mathbf{50}}$	${\mathbf{AP}}_{\mathit{S}}$
VisDrone	–	–	0.3010	0.5053	0.2202
	–	✓	0.3087	0.5171	0.2275
	✓	–	0.3225	0.5426	0.2493
	✓	✓	0.3280	0.5483	0.2514
UAVDT	–	–	0.2376	0.4033	0.2346
	–	✓	0.2545	0.4463	0.2570
	✓	–	0.2271	0.3928	0.2685
	✓	✓	0.2466	0.4153	0.3273

: the stride-4 level supplies candidate locations, and the assignment rule decides which of them can become confident predictions. Neither effect is redundant with the other, and the gain from both shrinks monotonically with object size ($+11.1$ recall points below 16 pixels, $+1.9$ above 64).

4.8. Full-Factorial Ablation of Assignment and P2 Refinement

To separate the main effects of high-resolution refinement and assignment restriction, we train a full-factorial set under a fixed uniform-loss recipe: the common detector, $P_2$-only, CGA-only, and CGA + $P_2$. Input size, optimizer, learning rate, training steps, augmentation, and checkpoint selection are fixed within each dataset. All rows are evaluated at the original image resolution.

Table 10 shows a consistent VisDrone pattern. Adding only the $P_2$ refinement branch improves the common detector by $0.0077$ mAP and $0.0073$ ${\mathrm{AP}}_S$; adding CGA alone gives a larger gain of $0.0215$ mAP and $0.0291$ ${\mathrm{AP}}_S$; combining both reaches $0.3280$ mAP and $0.2514$ ${\mathrm{AP}}_S$. Thus, the dense multi-class UAV setting benefits from both a scale-constrained positive set and localized stride-4 feature capacity.

The UAVDT factorization exposes a different transfer trade-off. $P_2$ alone gives the highest mAP ($0.2545$), whereas CGA + $P_2$ gives the highest ${\mathrm{AP}}_S$ ($0.3273$), improving over the common detector by $9.27$ ${\mathrm{AP}}_S$ points. Thus, the complete pathway changes the operating point toward extremely small vehicles, while aggregate mAP favors a less constrained variant. This is not a contradiction: it shows that the value of the pathway is best understood through the target metric and scene regime, rather than through mAP alone.

4.9. Is the Stride-4 Level Worth Its Cost?

Table 10 varies the refinement branch, but the stride-4 level itself is present in all of its rows. Because that level is the single most expensive decision in the design, we ablate it directly. The $P_3$–$P_5$ variant keeps the backbone, neck design, head, optimizer, schedule, augmentation, seed and evaluation protocol of the factorial recipe, and changes only the pyramid depth; the overlapping-range policy is preserved, so each level still starts at half the upper bound of the previous one. Accuracy is measured at the original resolution and latency on an otherwise idle GPU.

Table 11. Direct ablation of the stride-4 level on VisDrone. Backbone, neck design, head, schedule, augmentation, seed, and evaluation protocol are identical across rows; only the pyramid depth changes. Latency is end-to-end at ${1024}^2$ with batch size 1 on an idle RTX 4090 D.

Pyramid	mAP	${\mathbf{AP}}_{\mathbf{50}}$	${\mathbf{AP}}_{\mathit{S}}$	Params (M)	Locations	Latency (ms)	Memory (MB)
$P_3$–$P_5$ + CGA	0.2927	0.4898	0.2003	38.83	21,504	22.34	456
$P_2$–$P_5$ + CGA	0.3225	0.5426	0.2493	41.81	87,040	45.88	811
$P_2$–$P_5$ + CGA + $P_2$ refine	0.3280	0.5483	0.2514	42.99	87,040	49.76	815

separates two decisions that Table 10 conflates. Adding the stride-4 level is worth $2.98$ mAP and $4.90$ ${\mathrm{AP}}_S$ and costs $2.05\times$ the end-to-end latency, four times the prediction locations and $1.78\times$ the peak memory. Adding the refinement branch on top is worth a further $0.55$ mAP for $1.18$ M parameters and $8.5\%$ more latency. The expensive decision is therefore also the decisive one, and the cheap one is genuinely marginal.

UAVDT qualifies this conclusion rather than confirming it. There, the stride-4 level buys $10.4$ points of ${\mathrm{AP}}_S$ ($0.2230\to 0.3273$) but only $0.71$ points of mAP, and ${\mathrm{AP}}_{50}$ actually falls by $1.2$ points. The level is thus worth its cost when the deployment target is recall of extremely small objects; when the target is aggregate mAP, or when compute is constrained, $P_3$–$P_5$ is the better operating point.

4.10. Sensitivity to Center Radius, P2 Depth, and Regression Range

We next ask how sharply the pathway depends on its hyperparameters. The reference configuration uses two $P_2$ refinement blocks, center radius $r=1.5$, and compact overlapping regression ranges $[0,32]$, $[16,64]$, $[32,128]$, and $[64,+\infty )$.

Table 12. VisDrone sensitivity of $P_2$ depth, center radius, and regression ranges. All rows use the CGA + $P_2$ pathway and original-resolution evaluation.

Setting	mAP	${\mathbf{AP}}_{\mathbf{50}}$	${\mathbf{AP}}_{\mathit{S}}$
Reference ($r=1.5$, 2 blocks, default ranges)	0.3280	0.5483	0.2514
$P_2$ blocks = 1	0.3272	0.5471	0.2472
$P_2$ blocks = 4	0.3267	0.5422	0.2439
Center radius $r=1.0$	0.3342	0.5569	0.2531
Center radius $r=2.0$	0.3236	0.5403	0.2441
No-overlap ranges	0.3235	0.5390	0.2452
Wide overlapping ranges	0.3308	0.5515	0.2543

reports single-run VisDrone sensitivity under the same original-resolution protocol.

The $P_2$ depth sweep places the two-block reference near a practical optimum: one and four blocks remain close in mAP, but both lose ${\mathrm{AP}}_S$. The center-radius sweep provides the clearer explanation. Tightening the center neighborhood to $r=1.0$ raises VisDrone mAP to $0.3342$, whereas widening it to $r=2.0$ lowers both mAP and ${\mathrm{AP}}_S$. The result is consistent with the idea that additional stride-4 locations are useful only when ambiguous off-center positives are filtered. A single run cannot separate a real effect from initialization noise, so we repeated both settings with three seeds under the same recipe (

Table 13. Center radius under three random seeds (42, 123, 456), VisDrone validation split at original resolution, factorial recipe. Entries are mean±population standard deviation.

Setting	mAP	${\mathbf{AP}}_{\mathit{S}}$	Individual Runs (mAP)
$r=1.5$	$0.3254\pm 0.0021$	$0.2469\pm 0.0047$	0.3280/0.3252/0.3229
$r=1.0$	$\mathbf{0.3329}\pm \mathbf{0.0003}$	$\mathbf{0.2537}\pm \mathbf{0.0002}$	0.3333/0.3327/0.3327

). On VisDrone, the effect is real: $r=1.0$ gains $0.75$ mAP and $0.68$ ${\mathrm{AP}}_S$ against a within-setting dispersion of $0.0003$ and $0.0021$, roughly six times the run-to-run spread. The direction agrees with the assignment audit of Section 4.7: because $r{s}_l$ is an absolute distance, $r=1.5$ corresponds to $\pm 6$ pixels at $P_2$, which is permissive for objects only a dozen pixels across.

That alone would argue for changing the default, so we tested whether it transfers.

Table 14. Does the $r=1.0$ advantage transfer? The full main configuration is retrained at $r=1.0$ on every dataset with three seeds and evaluated on both splits, against the reported $r=1.5$ results of Table 3 and Section 4.12. Positive values favor $r=1.0$.

Dataset	Split	mAP ($\mathit{r}=1.5$)	mAP ($\mathit{r}=1.0$)	$\mathbf{\Delta }$ mAP	$\mathbf{\Delta }{\mathbf{AP}}_{\mathit{S}}$
VisDrone	validation	0.326	0.332	$+0.56$	$+0.09$
UAVDT	validation	0.261	0.261	$-0.00$	$+0.70$
UAVVaste	validation	0.488	0.488	$-0.02$	$-0.99$
VisDrone	test	0.244	0.248	$+0.43$	$+0.25$
UAVDT	test	0.163	0.156	$-0.74$	$-0.56$
UAVVaste	test	0.528	0.522	$-0.63$	$+0.15$
Mean				$-0.07$	$-0.06$

retrains the full main configuration at $r=1.0$ on all three datasets with three seeds each, on both splits. It does not transfer: the VisDrone gain persists, but UAVDT and UAVVaste are unchanged on validation and lose ground on test, and the six-combination mean is $-0.07$ mAP. Selection rule (ii) of Section 4.2 forbids adopting a value that is favorable on one dataset and neutral-to-harmful on the others, since that would encode a VisDrone-specific preference into a setting presented as general. We therefore retain $r=1.5$ and report $r=1.0$ as a documented option for dense multi-class scenes.

This also bounds how far the sensitivity study in Table 12 should be read. It is a single-dataset sweep, and the entries that look best there are not guaranteed to survive on another benchmark; the same caution applies to the wider regression ranges, whose VisDrone advantage likewise reverses on UAVDT. Regression-range overlap behaves similarly: removing overlap is harmful, while wider overlap recovers the best ${\mathrm{AP}}_S$ ($0.2543$) but does not surpass the $r=1.0$ setting in mAP.

Table 15. UAVDT regression-range sensitivity for the CGA + $P_2$ pathway under original-resolution evaluation.

Range Setting	mAP	${\mathbf{AP}}_{\mathbf{50}}$	${\mathbf{AP}}_{\mathit{S}}$
Default overlap	0.2466	0.4153	0.3273
No-overlap	0.2223	0.3648	0.2540
Wide overlap	0.2512	0.4121	0.3154

shows regression-range sensitivity on UAVDT, where the validation set is dominated by small vehicles. The no-overlap configuration is clearly harmful, indicating that adjacent levels need a shared transition region when object scales are close to the feature-stride boundary. A wider overlap improves mAP relative to the default, but the compact default retains the best ${\mathrm{AP}}_S$. This mirrors the factorial result: overall ranking and extremely small-object ranking select different operating points.

4.11. Quality and Head Variants of the Scale-Isolated Pathway

We also evaluated two quality-oriented variants under the same protocol: a soft centerness-quality classification target [34,35], and a scale-adaptive Wasserstein-IoU localization term for tiny boxes.

Table 16. Single-run pathway variants evaluated at original resolution (mAP/${\mathrm{AP}}_S$).

Variant	VisDrone	UAVDT	UAVVaste
CPD-FCOS	0.327/0.249	0.272/0.288	0.495/0.196
+soft-quality cls.	0.311/0.228	0.258/0.295	0.491/0.187
+SA-WIoU loc.	0.328/0.251	0.236/0.242	0.486/0.173

shows a useful separation: soft-quality classification improves UAVDT ${\mathrm{AP}}_S$ but costs VisDrone accuracy, while SA-WIoU gives the strongest VisDrone variant but transfers poorly to UAVDT. The plain configuration is therefore not an arbitrary simplification but the most balanced choice across the three regimes, and the variants indicate where score calibration and localization modeling could be specialized in future work.

We also tested a scale-decoupled head with separate classification and regression towers at the finest pyramid level. On UAVDT, this dedicated head improves over the common detector but does not outperform the simpler $P_2$ refinement pathway. The comparison suggests that the main leverage comes from improving the representation before prediction, rather than duplicating prediction towers without changing the underlying feature quality.

4.12. Independent Ablation of Scale-Aware Loss Reweighting

Our final model uses the uniform positive-sample loss of Equation (5). To justify this design choice, this subsection examines whether additionally introducing area-based scale-aware reweighting on top of the final model (i.e., $w_j=\mathrm{clip}({(s_0^2/A_j)}^{\gamma },1,w_{\max })$ with $s_0$ = 48, $\gamma$ = 0.5, $w_{\max }$ = 3) is beneficial. The two configurations keep all other training settings and recipes identical and are trained with three random seeds and evaluated on the complete validation split at the original resolution, reporting mean and standard deviation (

Table 17. Independent ablation of scale-aware loss reweighting, three-seed mean ± std (original resolution, complete validation split, torchmetrics). “Ours” is the final model (uniform positive-sample weights); “+reweighting” adds scale-aware weights on top of it.

Dataset	Config	mAP	${\mathbf{AP}}_{\mathit{S}}$
VisDrone	Ours	$\mathbf{0.326}\pm \mathbf{0.001}$	$\mathbf{0.251}\pm \mathbf{0.002}$
	+reweighting	$0.319\pm 0.006$	$0.245\pm 0.003$
UAVDT	Ours	$0.261\pm 0.010$	$\mathbf{0.287}\pm \mathbf{0.024}$
	+reweighting	$0.261\pm 0.016$	$0.264\pm 0.023$
UAVVaste	Ours	$\mathbf{0.488}\pm \mathbf{0.007}$	$\mathbf{0.196}\pm \mathbf{0.012}$
	+reweighting	$0.487\pm 0.004$	$0.189\pm 0.011$

). The three-seed comparison gives a consistent answer: area reweighting leaves mAP roughly unchanged, but lowers ${\mathrm{AP}}_S$ on all three datasets by $0.6$, $2.3$, and $0.7$ points on VisDrone, UAVDT, and UAVVaste, respectively. Thus, the small-object gains do not come from a generic preference for small boxes; they are tied to the pathway’s feature and positive-location design. Uniform positive-sample weights are consequently retained in the final model.

Run-to-run stability is reported by the “Ours” rows of Table 17: the VisDrone advantage over the strongest single-run baseline is much larger than the inter-seed dispersion, while UAVDT shows the widest spread, consistent with a more seed-sensitive vehicle-only transfer regime. Three seeds do not constitute a formal hypothesis test, but they do show that the principal VisDrone pattern is not tied to a single initialization.

4.13. Efficiency Analysis

Table 18. Efficiency comparison (single $1024\times 1024$ input, batch size 1, RTX 4090 D). GFLOPs are framework-specific profiler estimates and exclude dynamic detection post-processing. Latency is implementation-specific and is not a controlled cross-framework or end-to-end onboard comparison. Each model is measured after clearing previously resident models from GPU memory.

Method	Params (M)	GFLOPs	Latency (ms)	FPS	Memory (MB)
Faster R-CNN	41.3	210.6	19.8	50.4	657
RetinaNet	32.4	211.8	17.6	56.9	430
FCOS	32.1	206.1	18.4	54.3	429
YOLOv8s	11.2	73.8	4.2	237.8	216
YOLOv11s	9.5	55.6	6.0	166.2	235
CPD-FCOS (Ours)	43.0	766.6	46.4	21.5	724

reports parameters, GFLOPs (excluding post-processing), implementation-specific forward latency, throughput, and peak memory for our method and representative detectors on a single $1024\times 1024$ image with batch size 1 on the same RTX 4090 D GPU. CPD-FCOS requires about 3–$4\times$ the FLOPs of the ResNet-50 FPN-based detectors, mainly because the shared head also evaluates the dense stride-4 $P_2$ map. The resulting $21.5$ FPS and higher peak memory make the trade-off visible: the pathway spends substantial computation to retain fine-scale locations, whereas the YOLO models provide a much smaller compute envelope. The latency values are framework-specific and should not be read as a controlled cross-framework ranking, but they clarify the intended use of the method.

Table 18 times the network forward only, which understates deployment cost.

Table 19. End-to-end latency by stage (single $1024\times 1024$ image, batch size 1, idle RTX 4090 D, median over 50 real validation images). The $P_3$–$P_5$ column is the pyramid ablation of Table 11.

Stage	CPD-FCOS (ms)	${\mathit{P}}_3$–${\mathit{P}}_5$ (ms)	Ratio
Preprocessing	2.28	2.02	1.13×
Backbone + neck	22.73	12.24	1.86×
Detection head	21.53	6.15	3.50×
Decoding + NMS	2.93	1.76	1.67×
Remap to original resolution	0.16	0.11	1.45×
End-to-end	49.76 (20.1 FPS)	22.34 (44.8 FPS)	2.23×

therefore breaks the full pipeline into five stages, measured on 50 real validation images after 20 warm-up iterations on an otherwise idle GPU; real images are used because the cost of decoding and non-maximum suppression depends on how many candidates survive thresholding. End-to-end latency is $49.76$ ms, or $20.1$ FPS.

The breakdown corrects an expectation we held when designing the study. A stride-4 level enlarges the candidate grid fourfold, so we expected decoding and NMS to dominate; they grow, but account for only $5.9\%$ of end-to-end latency. The cost falls almost entirely on the shared head, which evaluates four convolutional towers at $\mathrm{87,040}$ locations, is $3.50\times$ slower than its $P_3$–$P_5$ counterpart and alone accounts for $43\%$ of the total. Efficiency work should therefore sparsify head computation at high-resolution levels—as sparse-query detectors [8] do—rather than optimize post-processing.

5. Discussion

5.1. Mechanistic Interpretation and Cross-Dataset Behavior

The full-factorial ablation supports the central design rationale in dense VisDrone scenes: both the scale-constrained positive set and localized $P_2$ refinement contribute measurable gains, and their combination gives the strongest result. This complementarity is the key evidence that CPD-FCOS is more than a high-resolution head appended to a generic detector. Section 4.7 identifies where that complementarity comes from, and it is not where we first assumed. The refinement branch is not a selective edge enhancer: its residual is comparable in magnitude to the base feature and only $1.23$ times stronger inside ground-truth boxes than outside. What it supplies is capacity, spread broadly—which is why it raises small-object recall while leaving AP unchanged, the extra detections being matched by extra false positives, and why it degrades large-object precision. Constrained assignment supplies the missing selectivity. Representation and supervision are thus complementary in a measurable sense: one provides candidate locations, the other decides which can be trusted.

The sensitivity study further shows that the pathway is not a single-point artifact. Moderate changes to $P_2$ depth remain close to the reference result, and the three-seed comparison of Table 13 establishes that a tighter center window ($r=1.0$) is a genuine improvement on VisDrone rather than seed noise. Table 14 then shows that this particular gain does not transfer, which is a useful negative result: a hyperparameter tuned on the densest benchmark can be neutral or harmful elsewhere, and single-dataset sweeps should not be promoted to defaults.

The locked tests qualify that validation-based interpretation. VisDrone retains a clear advantage over the validation-selected YOLOv11s reference. UAVDT does not show a clear overall advantage over FCOS and has lower small-object AP, despite the strong validation ${\mathrm{AP}}_S$ of the complete pathway. The UAVVaste mAP difference is modest, and its apparently larger ${\mathrm{AP}}_S$ gain is based on a few small annotations. Hence, if a deployment prioritizes extremely small-object recall, checkpoints should not be selected by mAP alone; ${\mathrm{AP}}_S$, recall, false-positive tolerance, and target-domain validation all require monitoring.

The validation–test change is not unique to CPD-FCOS. On VisDrone, CPD-FCOS changes from 0.326 validation mAP to 0.244 test mAP, while YOLOv11s changes from $0.288\pm 0.004$ to $0.223\pm 0.001$; the margin consequently narrows from 3.8 to 2.1 points. On UAVDT, CPD-FCOS changes from 0.261 to 0.163 and FCOS from $0.224$ to $0.167$. On UAVVaste, both CPD-FCOS and YOLOv8s increase on test, from 0.488 to 0.528 and from $0.482\pm 0.003$ to $0.517\pm 0.005$, respectively. These parallel movements are consistent with a split effect, although the non-paired training recipes preclude a causal claim.

A post hoc split audit helps characterize, but does not causally explain, these shifts. VisDrone contains 70.73 objects/image in validation versus 46.65 in test, with small-object fractions of 68.6% and 67.7%. UAVDT changes more substantially: the small-object fraction rises from 67.1% in validation to 74.9% in test and median box area falls from 696 to 496 pixels². UAVVaste test contains 507 boxes, only 6.3% of which are small, compared with 12.5% of 616 validation boxes. These descriptive differences, together with video-scene dependence and the absence of repeated baseline seeds, limit causal attribution of the validation–test gaps. They nevertheless show why validation evidence must not be presented as final cross-domain performance.

5.2. Deployment Boundaries and Residual Risks

The $P_2$ refinement branch adds only about $2.8\%$ parameters and preserves a single-forward inference flow, making the accuracy gain structurally targeted rather than the result of uniformly widening the detector. The full stride-4 prediction path nevertheless raises total computation to roughly 3–$4\times$ that of ResNet-50-FPN-type detectors. CPD-FCOS therefore makes a deliberate accuracy–efficiency trade-off: it is a strong desktop-GPU reference for dense fine-scale detection, while deployment on power-limited airborne hardware calls for sparse or dynamic high-resolution execution.

Several limitations remain. The strongest baseline of each dataset is now a three-seed mean, but the remaining comparison detectors are still single runs, so Table 5 should be read as a system-level reference rather than a statistical test; the design claims rest on the controlled experiments instead, where every factor except one is held fixed. The current input pipeline uses square resizing; an aspect-ratio-preserving pipeline should be tested before deployment in applications where geometric distortion is critical. The UAVDT test result shows appreciable inter-seed variance, and the small UAVVaste test split limits interpretation of sub-percentage differences. The primary VisDrone conversion excludes ignored annotations but does not suppress detections overlapping ignored regions, so the locked COCO-style values are not official challenge scores. The post hoc official-tool validation audit preserves the CPD-FCOS–YOLOv11s ordering, but it neither replaces a locked official-tool test nor supplies area-partitioned AP. Finally, all timings in this paper are desktop-GPU measurements. We did not have access to an embedded airborne platform, so onboard latency and energy consumption remain unmeasured and are not estimated here; they must be characterized before deployment. The end-to-end breakdown does, however, indicate where such an effort should start, since the shared head at the stride-4 level rather than post-processing dominates the cost.

6. Conclusions

We presented CPD-FCOS, a single-stage UAV detector in which a real stride-4 $P_2$ level, a localized residual refinement branch and center-constrained positive selection are designed as one pathway, and we measured what each part contributes. Direct ablation shows that the stride-4 level accounts for most of the accuracy gain and for essentially all of the added computation, while the refinement branch is cheap and marginal. Resolving the results by object size shows why the two are complementary: added capacity alone raises small-object recall without raising average precision, and only constrained supervision converts it into precision. The same analysis delimits the claim: the advantage grows as objects shrink and reverses on the largest ones; it is clear in dense multi-class scenes, becomes an operating-point shift on vehicle-only footage, and is not measurable where small annotations are few. CPD-FCOS is therefore offered as a controlled and reproducible account of how fine-scale representation and fine-scale supervision interact, and of what that interaction costs. Detectors that add region cropping or repeated inference remain ahead on aggregate VisDrone accuracy; the pathway studied here is compatible with them, and we hope the measurements reported above are useful to anyone combining the two.