DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR

• DOI: 10.48550/arxiv.2201.12329
• Publication Date: 2022-01-01
• Authors: Shilong Liu; Feng Li; Hao Zhang; Xiao Yang; Xianbiao Qi; Hang Su; Jun Zhu; Lei Zhang
• Keywords: Pooling; Benchmark (surveying); Geocoding

We present in this paper a novel query formulation using dynamic anchor boxes for DETR (DEtection TRansformer) and offer deeper understanding of the role queries DETR. This new directly uses box coordinates as Transformer decoders dynamically updates them layer-by-layer. Using not only helps explicit positional priors to improve query-to-feature similarity eliminate slow training convergence issue DETR, but also allows us modulate attention map width height information. Such design makes it clear that can be implemented performing soft ROI pooling layer-by-layer cascade manner. As result, leads best performance on MS-COCO benchmark among DETR-like detection models under same setting, e.g., AP 45.7\% ResNet50-DC5 backbone trained 50 epochs. conducted extensive experiments confirm our analysis verify effectiveness methods. Code is available at \url{https://github.com/SlongLiu/DAB-DETR}.