Equalization Loss for Large Vocabulary Instance Segmentation

TL;DR

This paper introduces an equalization loss to address the long tail of rare categories in large vocabulary instance segmentation, significantly improving performance on the LVIS dataset.

Contribution

The paper proposes a novel equalization loss method that effectively tackles the rare categories problem in large vocabulary datasets, enhancing detection and segmentation accuracy.

Findings

Achieves 5.1% overall AP gain on LVIS

Improves AP of rare categories by 11.4%

Ranks 1st in LVIS Challenge 2019

Abstract

Recent object detection and instance segmentation tasks mainly focus on datasets with a relatively small set of categories, e.g. Pascal VOC with 20 classes and COCO with 80 classes. The new large vocabulary dataset LVIS brings new challenges to conventional methods. In this work, we propose an equalization loss to solve the long tail of rare categories problem. Combined with exploiting the data from detection datasets to alleviate the effect of missing-annotation problems during the training, our method achieves 5.1\% overall AP gain and 11.4\% AP gain of rare categories on LVIS benchmark without any bells and whistles compared to Mask R-CNN baseline. Finally we achieve 28.9 mask AP on the test-set of the LVIS and rank 1st place in LVIS Challenge 2019.