Qi Tianhao ; Xie Hongtao ; Li Pandeng ; Ge Jiannan ; Zhang Yongdong

Abstract

Conventional detectors suffer from performance degradation when dealing withlong-tailed data due to a classification bias towards the majority headcategories. In this paper, we contend that the learning bias originates fromtwo factors: 1) the unequal competition arising from the imbalanceddistribution of foreground categories, and 2) the lack of sample diversity intail categories. To tackle these issues, we introduce a unified frameworkcalled BAlanced CLassification (BACL), which enables adaptive rectification ofinequalities caused by disparities in category distribution and dynamicintensification of sample diversities in a synchronized manner. Specifically, anovel foreground classification balance loss (FCBL) is developed to amelioratethe domination of head categories and shift attention todifficult-to-differentiate categories by introducing pairwise class-awaremargins and auto-adjusted weight terms, respectively. This loss prevents theover-suppression of tail categories in the context of unequal competition.Moreover, we propose a dynamic feature hallucination module (FHM), whichenhances the representation of tail categories in the feature space bysynthesizing hallucinated samples to introduce additional data variances. Inthis divide-and-conquer approach, BACL sets a new state-of-the-art on thechallenging LVIS benchmark with a decoupled training pipeline, surpassingvanilla Faster R-CNN with ResNet-50-FPN by 5.8% AP and 16.1% AP for overall andtail categories. Extensive experiments demonstrate that BACL consistentlyachieves performance improvements across various datasets with differentbackbones and architectures. Code and models are available athttps://github.com/Tianhao-Qi/BACL.