Reproduce “Semi-Supervised Multi-View Crowd Counting by Ranking Multi-View Fusion Models” 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

Qi Zhang, Yunfei Gong, Zhidan Xie, Zhizi Wang, Antoni B. Chan, Hui Huang

Semi-supervised multi-view crowd counting is formulated around consistency relations between predictions from different camera subsets, allowing synchronized but unlabeled frames to contribute training signal. The proposed MVPR and MVUR methods reduce annotation requirements across synthetic and real datasets, with MVUR achieving MAEs of 9.73/9.01/8.09 on CVCS, 7.79/6.51/5.82 on CityStreet, and 3.68/3.60/3.33 on PETS2009 using only 5%/10%/20% labeled data.

Key insights

  • Different camera configurations provide a source of weak supervision: predictions from fewer and more views can be compared even when no ground-truth density map is available.
  • Uncertainty ranking is more robust than count ranking when camera views are redundant, as observed on CityStreet, where extra views improve confidence without necessarily increasing estimated crowd size.
  • The approach uses unlabeled synchronized images without pseudo-label generation, but still depends on calibrated cameras, reliable scene projection, and assumptions about how additional views affect predictions.

Summary by alphaXiv.

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