Balanced Activation for Long-tailed Visual Recognition
This work addresses the mismatch between training and testing distributions in object detection for real-world long-tailed data, offering a simple and effective solution.
The paper tackles the problem of long-tailed visual recognition by introducing Balanced Activation, an extension of Sigmoid and Softmax functions to address label distribution shifts, resulting in a ~3% mAP gain on LVIS-1.0 and outperforming state-of-the-art methods without extra parameters.
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 report, we introduce Balanced Activation (Balanced Softmax and Balanced Sigmoid), an elegant unbiased, and simple extension of Sigmoid and Softmax activation function, to accommodate the label distribution shift between training and testing in object detection. We derive the generalization bound for multiclass Softmax regression and show our loss minimizes the bound. In our experiments, we demonstrate that Balanced Activation generally provides ~3% gain in terms of mAP on LVIS-1.0 and outperforms the current state-of-the-art methods without introducing any extra parameters.