Underwater Object Detection Based on DN-DETR

We present a method for underwater object detection with DN-DETR (DeNoising-DEtection TRansformer) and DN-Deformable-DETR. Both DN-DETR and DN-Deformable-DETR are DETR-like detectors that introduce a denoising acceleration training method. Denoising is capable of accelerating the convergence rate and reducing the training instability of the native DETR detector. In addition, DETR has a strong ability on learning global image context and performs better in the detection of large and small targets. The results show that DN-Deformable-DETR has better average precision (+3.7AP ) and better performance on large targets (+5.1AP L) and small targets (+11.5AP S) than Faster R-CNN. Compared with DN-Deformable-DETR, DN-DETR obtained the optimal value in a single metric (+0.8AP 50).