VDDT: Improving Vessel Detection with Deformable Transfomer

원문 초록

Vessel detection has received wide attention in object detection, and the recently proposed DETR has successfully achieved true end-to-end object detection and has shown good performance. However, DETR is not sensitive to detect small objects, resulting in its unsatisfactory performance in vessel detection. In this paper, we use Deformable DETR as the baseline model and modify it on top of that. Firstly, we add reference point information to object queries to make the features learned by object queries richer to improve the performance of the detector. Secondly, we use multi-layer perceptron instead of multi-head self-attention to reduce the computational effort of the decoder. In addition, we collected 85 videos annotated with 4563 images and used these images to make a vessel dataset. The experimental data on our vessel dataset shows that VDDT performs better compared to the baseline.