PaceShift adds guarded curriculum training for dense prediction
Original title: Supervision-Adaptive Curriculum and Guarded Anti-Curriculum Learning for Dense Prediction
The method keeps the deployed architecture unchanged and adds no inference computation; code is not yet public.
Researchers present PaceShift, extending CurriSeg with a stabilize-then-challenge approach for dense prediction. Phase one weights samples by supervision reliability and pixel uncertainty; phase two applies guarded spectral challenges against a frozen reference. The framework covers full, weak, semi-supervised, and model-derived supervision. Reported experiments show accuracy gains plus better pooled calibration and corruption robustness in the tested settings.
- Reliability uses supervision, augmentation consistency, disagreement, and prediction certainty.
- A pairwise preference objective is measured against the frozen Phase-I reference.
- Evaluations include CECS, broader dense prediction, and foundation-model adaptations.
- The paper says code will be publicly released.
From the original article
1 min read at arxiv.orgCurriculum learning orders samples by difficulty, but difficulty alone does not establish supervision reliability for dense prediction. Easy samples can expose shortcuts, hard samples can be informative, and noisy or ambiguous samples can mislead training.
By Chunming He · 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.