Patch-Based Auxiliary Node Classification for Domain Adaptive Object Detection

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper addresses the challenge in domain adaptive object detection (DAOD), which is to train object detection models on labeled source domain data that can generalize to unlabeled target domains. Traditional methods have used graph representations for fine-grained pixel-level domain adaptation, but they have faced issues with suboptimal graph matching due to node class confusion. This paper introduces a patch-based auxiliary node classification method for DAOD that enriches node context information by utilizing local region information of nodes and employs convolutional neural networks to learn local region feature representation. This improves node classification accuracy and reduces the risk of class confusion. The paper also proposes a progressive strategy for fusing inherent features with learned local region features for stable and reliable classification. Experiments on various DAOD scenarios show that this model outperforms existing works.

1. Line 46-50: The introduction outlines the issue of domain shift in object detection and the use of unsupervised domain adaptation (UDA) methods. It may be beneficial to include recent examples of practical applications where UDA has effectively addressed domain shift, to provide concrete evidence of its relevance in real-world scenarios.

2. Line 149-160: In the proposed patch-based auxiliary node classification method, consider exploring the impact of different sizes and shapes of patches on the classification performance. This could help in understanding how much local context is beneficial and could lead to further optimization of the model.

3. Line 250-300 : The progressive node feature fusion strategy is introduced, which is crucial for the model's performance. It could be advantageous to discuss the potential risks or challenges associated with this strategy, such as overfitting to the training data or difficulties in parameter tuning.

4. Line 381-386: The paper describes the experimental setup and the use of FCOS and VGG-16. Given the rapid development of object detection architectures, it may be valuable to compare the performance of the proposed method with other state-of-the-art architectures to ensure the model's adaptability and robustness.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper is well written, provides well organized and motivated theoretical part along with enough and convincing experimental results, which are compared with the results of recent methods in the field. Some technical and scientific notes that will help improve the presentation of the method are given below.

11.     On line 186 it is better to replace “illusory” with “imaginary” ;

32.    Line 196- the notion “fine-grained domain” needs a short definition;

43.   Figure 1 would need a short description. Otherwise the reader must read paper [11] as well, because this figure “illustrates the process of the proposed model”.

54.    The authors list their contributions in the Introduction. I would suggest they list them in a more concise form in the conclusion and tell where in the process shown in Figure 1 these contributions apply to?

65.  I believe \El_gothic-{node} and  \El_gothic-{n} are loss functions. If so tell this around Eqs. (1) and (3).

76.    As I see from Algorithm 1, the authors use, for their CNN, the loss function  \El_gothic-{n}. Please, briefly tell why do you use the latter loss function but not the former?

87.   I would suggest the authors prepare an overall block diagram, as Figure 1, in order to visually summarize the  “Progressive Node Feature Fusion Strategy” described in section 3.3.

98.   In Eq. (6) are used loss functions from [11] and [18]. Please clarify, is there in this paper any contribution regarding the loss function \El_gothic?

99.  I would suggest section Visualization Analysis e moved from Experiments to Conclusions section.

110. Shorten all long sentences like the one on lines 56-60.

Comments on the Quality of English Language

Accept after the authors address the notes given in their section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf