Host laboratory: Connaissance et Intelligence Artificielle Distribuées (CIAD) – http://www.ciad-lab.fr. Université de technologie de Belfort-Montbéliard (UTBM).
Keywords: deep learning, object re-identification (ReID), person re-identification (ReID), UAV vision, aerial imagery, computer vision, generative models (GAN, Stable Diffusion)
Contacts: Abderrazak Chahi (abderrazak.chahi@utbm.fr), Mohamed Kas (mohamed.kas@utbm.fr), Yassine Ruichek (yassine.ruichek@utbm.fr)
Description of the internship topic:
Object Re-Identification (ReID) is a major challenge in computer vision, particularly in identifying the same object (person, vehicle, animal, or infrastructure) across multiple images captured under varying conditions. In the context of unmanned aerial vehicles (UAVs), this task becomes even more complex and strategic. Drones operate in constant motion, often at high altitudes, with sudden changes in orientation, perspective, and scale. These extreme conditions drastically alter the visual characteristics of objects, making it difficult to learn robust representations that generalize to diverse environments [1]. Unlike fixed ground-based camera systems, UAVs introduce numerous geometric (rotations, tilts, oblique perspectives), photometric (lighting changes according to the drone’s orientation), and dynamic (speed of movement, vibrations, turbulence) variations. Additionally, object resolution may be low, viewing angles are rarely frontal, and scenes are often cluttered or moving, further increasing the difficulty of the problem [2,3].
Traditional state-of-the-art ReID approaches, often designed for stable urban environments, show their limitations when confronted with the variability of UAV aerial images. Existing deep models struggle to capture the extreme transformations produced by onboard cameras and do not consistently remain invariant to changes in orientation or altitude [4]. Simple data augmentation techniques are insufficient, as they do not accurately reproduce the diversity of disturbances encountered during flight. Similarly, conventional recognition architectures are not always suited to drastic changes in scale or to the appearance of dynamic occlusions caused by drone movements [5]. This situation highlights the need to develop new approaches capable of better modelling, simulating, or learning the transformations specific to aerial photography. In this context, this internship project aims to design, study, and compare various deep learning approaches to enhance the robustness of object/person re-identification (ReID) in aerial environments using unmanned aerial vehicles (UAVs). Potential research directions include developing modules that ensure invariance to extreme geometric transformations such as rotation, tilt, and perspective; designing advanced strategies for augmenting or generating realistic aerial data using generative models such as generative adversarial networks (GANs) and Stable Diffusion; investigating state-of-the-art architectures including visual transformers, multi-scale models, and hybrid networks; and implementing attention mechanisms that can isolate discriminative cues despite visual noise introduced by drone movement.
References:
[1] Chen, Shuoyi, et al. « Towards Effective Rotation Generalization in UAV Object Re-Identification. » IEEE Transactions on Information Forensics and Security (2025).
[2] Kim, Bongjun, et al. « Person Re-Identification with Attribute-Guided, Robust-to-Low-Resolution Drone Footage Considering Fog/Edge Computing. » Sensors (Basel, Switzerland) 25.6 (2025): 1819.
[3] Mei, Ling, et al. « Unsupervised Aerial-Ground Re-Identification from Pedestrian to Group for UAV-Based Surveillance. » Drones 9.4 (2025): 244.
[4] Albaluchi, Yousaf, et al. « UAV-based person re-identification: A survey of UAV datasets, approaches, and challenges. » Computer Vision and Image Understanding (2024): 104261.
[5] Nguyen, Huy, et al. « AG-VPReID: A Challenging Large-Scale Benchmark for Aerial-Ground Video-based Person Re-Identification. » Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.
Candidate Profile:
Holder or in the process of preparing a Master’s degree in computer science, computer vision, machine learning, robotics, or a related field.
- Advanced knowledge and practice in object-oriented programming (C++, Python) and machine learning tools (deep learning platforms: Pytorch, TensorFlow) are required.
- Advanced level in English writing and speaking is required.
Application (CV, scores, reference letters, …) to Abderrazak Chahi (abderrazak.chahi@utbm.fr), Mohamed Kas (mohamed.kas@utbm.fr), Yassine Ruichek (yassine.ruichek@utbm.fr) - Deadline: January 15 2026
