Communication Dans Un Congrès Année : 2024

Early feature distributions alignment in visible-to-thermal unsupervised domain adaptation for object detection

Résumé

Infrared or thermal images are used in many civilian and military applications to detect objects due to the heat they emit, especially when environmental conditions such as nighttime or adverse weather prevent the use of visible images. To train an object detector based on a deep neural network, a significant amount of annotated data is required to achieve good detection performance. However, annotations for infrared images are often unavailable and costly to obtain. Besides, the trained model may show poor robustness against the change of thermal sensor. Therefore, unsupervised domain adaptation (UDA) methods have been proposed to train an object detector with annotated visible images, which are easily available, and unannotated infrared images. This paper presents a new visible-to-thermal UDA approach for object detection based on Deformable-DETR with hybrid matching. Our approach aims to establish common features between visible and thermal images at the earliest stages of the backbone network. The feature distributions extracted from visible and thermal images are aligned thanks to discriminator networks and adversarial learning. Gradient images are also used as a domain translation of input images to ease the alignment. Detection performance is further improved by randomly masking tokens at the input of the transformer. Experiments on public datasets demonstrate that our method consistently outperforms previous works.
Fichier principal
Vignette du fichier
RGB_to_thermal_UDA_NoteSpringer.pdf (12.02 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

cea-04917972 , version 1 (28-01-2025)

Identifiants

Citer

Adrien Maglo, Romaric Audigier. Early feature distributions alignment in visible-to-thermal unsupervised domain adaptation for object detection. ICPR 2024 - 27th International Conference on Pattern Recognition, Dec 2024, Kolkata, India. pp.109-124, ⟨10.1007/978-3-031-78447-7_8⟩. ⟨cea-04917972⟩
0 Consultations
0 Téléchargements

Altmetric

Partager

More