Reproduce “SAGE: Saliency-Annealed Gated Encoder for Universal Vision Recognition” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Le Yang, Yucheng Huang, Junyu Chen, Qi Dang, Xin Kong, Haifeng Huo, Qiang Zhao

SAGE adapts pretrained Vision Transformers for dense vision, combining foreground and context cues with finer local features and task-specific decoders. Its schedule strengthens and withdraws enhancement modules while shifting from focal loss to Varifocal Loss. On COCO val2017, SAGE ViT-B reports 61.1 box AP and 54.4 mask AP, versus RF-DETR-2XL’s 60.1 and 49.9; architecture is not isolated.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.