PA-DETR : Prompt Adaptable DETR for object detection in unknown domains

<div> This work focuses on limiting the drop of performances in transfer settings without changing the model weights and without having access to large data from target domain on the complex task of object detection. The challenge lies in the fact that a domain change can damage representation of the searched objects, undesired object or even background in the latent space of the detection networks. Yet, by learning to slightly align those representations with given source prompts during training, we are able to leverage small target prompts at test time to adapt, without weight update, the model supervised latent representation. Throughout experiments on COCO, Objects365, Cityscapes and other datasets on state-of-the-art DETR detection head, we even prove that our method almost matches state-of-the-art domain adaptation architectures without weight update and with only 64 target domain labelled images for each given datasets. </div>