CARD: Certifiable Reweighting for Single Domain Generalization Object Detection

Fan Wu, Xue Zhao, Jinling Gao, Nanyang Ye

Keywords: single domain generalization, object detection

Abstract: Single Domain Generalization Object Detection (S-DGOD) is a challenging yet practical task, where we only have access to data from one specific source domain to train an object detection network, but have to generalize to numerous unseen target domains. Recent works point out that the learning dynamics of Deep Neural Networks (DNNs) are biased by gradient descent to learn simple semantics, which are usually non-causal and spuriously correlated to the ground truth labels, as a result, DNN-based object detection networks fail to consistently generalize well in the Out-of-Domain (OoD) scenario. In this paper, we focus on S-DGOD based on theoretical analysis, exploring a classic and widely-used approach, Generalizable Reweighting (GRW), which iteratively reweightes the training samples to improve generalization performance. In our theoretical analysis, we first identify that the vanilla GRW hardly outperforms Empirical Risk Minimization (ERM) in the S-DGOD scenario. To provide a generalization guarantee, we further derive Certifiable Feature Perturbation (CFP) based on our theory, which aims to train a robust object detection network against additional perturbations added to the extracted features. We demonstrate that GRW works well with CFP in achieving OoD generalization, thus, surpassing ERM by a large margin under worse conditions. This brand new reweighting strategy is named Certifiable Reweighting (CARD). Our extensive experiments show that the proposed CARD achieves SOTA performance compared to baseline methods on the five urban-scene S-DGOD benchmarks.

Primary Area: representation learning for 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/2024/AuthorGuide.

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: 2549