Dear Colleagues,

Object detection and recognition are central tasks in computer vision, which include the detection of objects boundaries and their classification. They have become essential in many applications, such as search and rescue, warehouse logistics, video surveillance or monitoring using UAVs, with low-resolution or blurred images usually captured due to camera motion. Additionally, the conditions may differ across different situations, making it complex to achieve general solutions; thus, fine-tuning is essential in new scenarios.

The computer vision community has adopted deep-learning models in the last decade due to their superior performance with respect to those from classical methods. These models require a high processing power (GPUs) for training with large datasets and provide inferences in real time; typically, these models employ convolutional neural networks (CNNs). They are subdivided in two types: two-shot detectors, that search with maximum accuracy with the potential cost of inference time; and one-shot detectors, which are oriented at a minimum inference time for real-time applications. Two-shot detectors are dominated by the R-CNN family (region-proposal CNNs), such as Fast R-CNN, Faster R-CNN or Cascade R-CNN solutions, while the YOLO family dominates one-shot detectors, being SSD and RetinaNet other popular algorithms in this category. Additionally, in recent years, Vision Transformers (ViTs) have also been applied to object detection and recognition tasks. ViT-based algorithms, such as DETR or YOLOS, are based on a self-attention mechanism that learns the relationships between elements of a sequence, applying the transformer architecture to image grids. ViTs make use of CNNs as a backbone for feature extraction, given their ability to automatically extract relevant features. In addition, object detection is closely related with other open challenges in machine vision such as Multi-Object Tracking (MOT), which involves both the detection and tracking of objects of interest appearing in the video sequence. The goal in this case is not only to identify and locate the objects contained in each frame, but to also associate them across frames to keep track continuity and follow their dynamics over time. This task is usually solved by combining algorithms addressing object detection and data association, and some relevant algorithms in the SORT family (Simple Online and Real-time Tracking) can be mentioned such as deepSORT, StrongSORT or OCT-Sort.

Regarding evaluation, developing fair comparisons among different solutions is complex, considering the balance between accuracy and speed, the resolution of the input images, the configuration of the evaluation parameters, etc. Analyses are based on the available benchmarks and datasets, which are necessary to evaluate the performance of different architectures and configurations. In this sense, many authors have identified class imbalance as an additional challenge to achieving a high accuracy. In this sense, other deep-learning architectures, such as GAN or autoencoders, can be combined with detectors to enhance the training phase, increasing the size and variety of the datasets, for instance, to improve the detection of very small objects. Additionally, learning can be improved for imbalanced situations, adapting the loss function to focus learning on hard examples and avoid a bias towards numerous negative examples.

This Special Issue is aimed at contributions focused on these topics, showing the capability of novel mathematical algorithms, architectures and methods to improve the object detection and recognition tasks, with the possibility of multi-object tracking, with an emphasis in new solutions and analysis of their performance in challenging conditions in relevant applications.

Prof. Dr. Jesús García-Herrero
Prof. Dr. Johan Debayle
Guest Editors

Manuscript Submission Information

Manuscripts should be submitted online at www.mdpi.com by registering and logging in to this website. Once you are registered, click here to go to the submission form. Manuscripts can be submitted until the deadline. All submissions that pass pre-check are peer-reviewed. Accepted papers will be published continuously in the journal (as soon as accepted) and will be listed together on the special issue website. Research articles, review articles as well as short communications are invited. For planned papers, a title and short abstract (about 100 words) can be sent to the Editorial Office for announcement on this website.

Submitted manuscripts should not have been published previously, nor be under consideration for publication elsewhere (except conference proceedings papers). All manuscripts are thoroughly refereed through a single-blind peer-review process. A guide for authors and other relevant information for submission of manuscripts is available on the Instructions for Authors page. Mathematics is an international peer-reviewed open access semimonthly journal published by MDPI.

Please visit the Instructions for Authors page before submitting a manuscript. The Article Processing Charge (APC) for publication in this open access journal is 2600 CHF (Swiss Francs). Submitted papers should be well formatted and use good English. Authors may use MDPI's English editing service prior to publication or during author revisions.

• object detection and classification
• multi-object tracking
• deep-learning architectures
• model training and evaluation
• loss functions in learning
• evaluation metrics and datasets
• class imbalance
• applications of object detection and object tracking
• aerial object identification
• transformers for object detection
• edge ai for object detection
• multi-modal object detection

Title: Transformers applied to object localization and classification in dense images
Authors: Gonzalo Rojas, Sergio Velastín, Jesús García
Affiliation: Universidad Carlos III de Madrid, Spain