Anchor DETR: Query Design for Transformer-Based Detector

In this paper, we propose a novel query design for the transformer-based object detection. previous detectors, queries are set of learned embeddings. However, each embedding does not have an explicit physical meaning and cannot explain where it will focus on. It is difficult to optimize as prediction slot specific mode. other words, on region. To solve these problems, in our design, based anchor points, which widely used CNN-based detectors. So focuses objects near point. Moreover, can predict multiple at one position difficulty: ``one region, objects''. addition, attention variant, reduce memory cost while achieving similar or better performance than standard DETR. Thanks proposed detector that called Anchor DETR, achieve run faster DETR with 10x fewer training epochs. For example, achieves 44.2 AP 19 FPS MSCOCO dataset when using ResNet50-DC5 feature 50 Extensive experiments benchmark prove effectiveness methods. Code available https://github.com/megvii-research/AnchorDETR.