Reproduce “Balancing Logit Variation for Long-tailed Semantic Segmentation” on OpenResearch

About this paper

Yuchao Wang, Jingjing Fei, Haochen Wang, Wei Li, Tianpeng Bao, Liwei Wu, Rui Zhao, Yujun Shen

Researchers from Shanghai Jiao Tong University, SenseTime Research, Institute of Automation (CAS), and CUHK developed Balancing Logit Variation (BLV), a plug-in module that addresses long-tailed data distributions in semantic segmentation by applying category-dependent stochastic perturbation to predicted logits. The method significantly boosts performance for under-represented classes and generalizes effectively across fully supervised, semi-supervised, and unsupervised domain adaptation settings.

Key insights

• Injecting category-dependent stochastic variation into logits during training can effectively 'expand' the feature space for tail classes, preventing feature squeezing and improving generalization.
• This approach generalizes robustly across diverse semantic segmentation tasks (fully supervised, semi-supervised, UDA) and various model architectures (CNNs and Transformers).
• The combination of random variation and a frequency-based balance coefficient is crucial for enriching training instances and achieving balanced performance without overfitting.

Summary by alphaXiv.