Margin Rectification for Long-Tailed Visual Recognition
Abstract: Long-tailed visual recognition tasks pose great challenges for neural networks on how to handle the imbalanced predictions between head and tail classes, i.e., models tend to classify tail classes as head classes. While existing research focused on data resampling and loss function engineering, in this paper, we take a different perspective: the \emph{classification margins}. We study the relationship between the margins and logits and empirically observe that the unrectified margins and logits are \emph{positively correlated}. We propose a simple yet effective \emph{MARgin Rectification} approach (\textbf{MARR}) to rectify the margins to obtain better logits. We validate MARR through extensive experiments on common long-tailed benchmarks including CIFAR-LT, ImageNet-LT, Places-LT, and iNaturalist-LT. Experimental results demonstrate that our MARR achieves favorable results on these benchmarks. In addition, MARR is extremely easy to implement with just three lines of code. We hope this simple approach will motivate people to rethink the unrectified margins and logits in long-tailed visual recognition.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: Mark the changes with blue texts:)
-----
1. We renamed the method from "calibration" to "rectification", which is more concise according to the comments of most reviewers.
2. We thoroughly re-examine all the minor weaknesses mentioned by reviewers and modify them.
3. We add more related work to discuss our contribution.
4. We resolve many questions and concerns raised by reviewers.
Assigned Action Editor: ~Yanwei_Fu2
Submission Number: 33
Loading