Differentiable Average Precision Loss in DETR
Keywords: Object detection，DEtection TRansformer, AP loss
TL;DR: we propose a novel loss function, Differentiable Average Precision Loss (DAP-loss), which provides a differentiable approximation of AP.
Abstract: Average Precision (AP) is a widely used metric for evaluating object detection systems because it effectively integrates both classification accuracy and localization precision. In this paper, we conduct a detailed analysis of the characteristics of the AP metric, focusing on its non-differentiability and non-convexity. Building on this analysis, we propose a novel loss function called Differentiable Average Precision Loss (DAP-loss), which provides a differentiable approximation of AP, thereby enabling direct optimization of AP across a set of images. We validate the effectiveness of DAP-loss both theoretically and empirically, extending its application to the cost functions used in the Hungarian matching algorithm, which makes it suitable for end-to-end detection models. DAP-loss supports the simultaneous optimization of classification and localization tasks within an end-to-end framework, eliminating the need for hyperparameters to balance these tasks—a common challenge in traditional methods. In the later stages of training, we applied DAP-loss to replace the original loss functions in several state-of-the-art end-to-end models, including DETR and Deformable DETR. Experimental results demonstrate that our method achieves significant improvements over baselines on the COCO dataset.
Primary Area: applications to computer vision, audio, language, and other modalities
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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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.
Submission Number: 2217
Loading