Reproduce “S$^3$FD: Single Shot Scale-invariant Face Detector” 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

Shifeng Zhang, Xiangyu Zhu, Zhen Lei, Hailin Shi, Xiaobo Wang, Stan Z. Li

S3FD develops a real-time, scale-invariant face detector that achieves state-of-the-art performance across all common benchmarks, notably improving detection of small faces by addressing limitations of anchor-based methods. The detector processes VGA-resolution images at 36 FPS on an Nvidia Titan X.

Key insights

  • Designing anchors based on Effective Receptive Fields (ERF) rather than theoretical receptive fields leads to better feature alignment and improved detection performance.
  • Continuously varying face scales require a more flexible anchor matching strategy beyond fixed Jaccard overlap thresholds to ensure sufficient positive samples for small and 'outer' faces.
  • Mitigating the extreme class imbalance in anchor-based detectors, especially for small anchors, is crucial, and approaches like max-out background labels can effectively reduce false positives.

Summary by alphaXiv.

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