Congratulations! VIPL's paper “Dual Compensation Residual Networks for Class Imbalanced Learning” is accepted by TPAMI. The full name of IEEE TPAMI is IEEE Transactions on Pattern Analysis and Machine Intelligence, which is an international journal on pattern recognition, computer vision and machine learning with an impact factor of 23.6, announced in 2023.
Ruibing Hou, Hong Chang, Bingpeng Ma, Shiguang Shan, Xilin Chen. Dual Compensation Residual Networks for Class Imbalanced Learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023.
An intuitive solution to address class imbalance problem is to re-balance data distribution, including re-sampling training data and re-weighting loss functions. However, the re-balanced distribution is easily fitted by the over-parameterized deep networks, increasing the risk of over-fitting the tail data. To address this problem, we propose a Feature Compensation Module (FCM) and Logit Compensation Module (LCM). The first module, FCM, alleviates overfitting on tail data by compensating for feature drift on tail classes. FCM is designed based on the following observation that the test tail features tend to drift towards the feature cloud of similar multiple head categories. To this end, FCM estimates a multi-mode feature drift direction for each tail category based on nearest head categories. The second module, LCM, translates the deterministic feature drift direction estimated by FCM along intra-class variations. In this way, the compensated features can cover a larger feature space and still live in true feature manifold, thus can better fit the test features. For evaluation, we conduct Class Imbalanced experiments on four long-tailed benchmarks. We demonstrate that various long-tailed methods can be directly incorporated into our proposed framework, yielding consistent performance improvements.
Download: