SRCD distills head-class CAMs to resist shortcuts on tail classes
Original title: Shortcut-Resistant CAM Distillation for Long-Tailed Recognition
The method aggregates a few head-class activation maps into a teacher and reports consistent gains on long-tailed benchmarks.
A paper argues that long-tailed training pushes models toward non-semantic shortcuts, especially on tail classes that share similar contexts. Shortcut-Resistant CAM Distillation aggregates class activation maps from a small set of head-class candidates with energy-model weights based on coherence and concentration, then distills that teacher to the tail-class map. The authors give a CAM-space analysis of shortcut-region evidence and say experiments improve strong baselines. Code is at github.com/Haifeng3/SRCD.
- Tail classes with few examples often share contexts, so background and texture cues correlate strongly.
- Head-class CAMs are weighted by coherence and concentration into a shortcut-resistant teacher.
- Theory treats shortcut reliance as shortcut-region evidence mass in CAM space.
- The framework is described as plug-and-play and is released with code.
From the original article
1 min read at arxiv.orgReal-world datasets often follow a long-tailed distribution, making generalization to tail classes difficult. We revisit this problem through the lens of shortcut learning, where models prefer the easiest predictive cues (e. g.
By Wenhai Wan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.