Class-Balanced Loss Based on Class Volume for Long-Tailed Object Recognition
ZhiJian Zheng (National University of Singapore),* Teck Khim Ng (National University of Singapore)
The 33rd British Machine Vision Conference
Abstract
The performance of classification neural networks is often suboptimal in the real world due to long-tailed data distributions. Re-sampling and re-weighting based on class frequency have been adopted in the literature to address the long-tailed problem. In this paper, we focus on the re-weighting approach. Re-weighting factors estimated by state-of-the-art approaches are determined by the number of samples which ignore the within class diversity (e.g. the cat class is visually more diverse than the frog class). In this paper, we propose a concept called class volume that measures the within class diversity and use this class volume to dynamically adjust the per-class weight. Our method does not introduce any hyperparameter and can be easily integrated into existing models with little computation overhead. We conducted extensive experiments and set the new state-of-the-art performance on widely-used long-tailed recognition benchmarks.Video
Citation
@inproceedings{Zheng_2022_BMVC,
author = {ZhiJian Zheng and Teck Khim Ng},
title = {Class-Balanced Loss Based on Class Volume for Long-Tailed Object Recognition},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0896.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by