Balanced Meta-Softmax for Long-Tailed Visual Recognition

TL;DR

This paper introduces Balanced Meta-Softmax, an unbiased extension of Softmax designed to address the bias caused by long-tailed data distributions in visual recognition, improving performance on classification and segmentation tasks.

Contribution

It proposes Balanced Softmax and Balanced Meta-Softmax, with theoretical analysis and a meta-sampling approach to enhance long-tailed visual recognition.

Findings

Outperforms state-of-the-art methods on visual recognition tasks

Provides theoretical generalization bounds for Softmax regression

Effectively handles label distribution shift in long-tailed data

Abstract

Deep classifiers have achieved great success in visual recognition. However, real-world data is long-tailed by nature, leading to the mismatch between training and testing distributions. In this paper, we show that the Softmax function, though used in most classification tasks, gives a biased gradient estimation under the long-tailed setup. This paper presents Balanced Softmax, an elegant unbiased extension of Softmax, to accommodate the label distribution shift between training and testing. Theoretically, we derive the generalization bound for multiclass Softmax regression and show our loss minimizes the bound. In addition, we introduce Balanced Meta-Softmax, applying a complementary Meta Sampler to estimate the optimal class sample rate and further improve long-tailed learning. In our experiments, we demonstrate that Balanced Meta-Softmax outperforms state-of-the-art long-tailed classification solutions on both visual recognition and instance segmentation tasks.