Akhat-DETR: End-to-End Object Detection Model on Hazy Scenarios in Autonomous Driving

The innovative DEtection TRansformer(DETR) approach introduces the transformer encoder and decoder architecture into object detection, obviating the need for hand-designed components. Though modern detectors have attained competitive results on public dataset such as COCO dataset, their capabilities are perverted on images captured in inclement weather. In this paper, we propose Akhat-DETR, an end-to-end transformer-based detector designed for hazy scenes. First, we design a light-weight convolutional dehazing network which can be integrated seamlessly into detectors. Moreover, we design a novel one-size-fits-all feature fusion module named FFTA. In the end, a general supervised learning design paradigm is given: as long as the final annotations are available, intermediate annotations are dispensable, thus the end-to-end model can perform training and inference in its entirety. Akhat-DETR achieves 61.0% AP on RTTS dataset with a 3090 GPU, triumphing over state-of-the-art detectors. Codes of proposed modules, splitted dataset in COCO format and pre-trained models are available at https://github.com/ChizkiyahuOhayon/Akhat-DETR.