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.
@article{arxiv.2008.11037,
title = {Balanced Activation for Long-tailed Visual Recognition},
author = {Jiawei Ren and Cunjun Yu and Zhongang Cai and Haiyu Zhao},
journal= {arXiv preprint arXiv:2008.11037},
year = {2020}
}
Comments
LVIS Challenge Workshop at ECCV 2020 Spotlight. arXiv admin note: substantial text overlap with arXiv:2007.10740