Guide Detectors in Pixel Space with Global Positioning and Abductive Matching

Yiyuan Zhang, Wencheng Han, Jianbing Shen

Keywords: Object Detection, Abductive DETR

Abstract: End-to-End object Detector ensembles prior knowledge in a concise framework. DETR (DEtection TRansformer) contains two steps: Learn object queries in the representation space and match the queries with boxes in the pixel space. The ambiguity of object queries in DETR lead to an uncertain assignment in the Hungarian Matching. The formulation loss in the pixel space will in turn affect the learning representations. Therefore, we propose the Abductive DETR, which learns object queries in the representation space with global positioning in the pixel space and matches object queries in the pixel space with the abductive awareness from the representation space. Experimentally, Abductive DETR can be transferred to other DETR-variants methods and achieves a satisfactory improvement. And it takes only 2 epochs to achieve the 98.7% accuracy of predicting the number of objects. Compared with other state-of-the-art methods on the MS COCO dataset, Abductive DETR also achieves outstanding performance and arrives at convergence much faster. Our code will be made publicly available soon.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)