Reproduce “Weakly Supervised Video Individual CountingWeakly Supervised Video Individual Counting” 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
Xinyan Liu, Guorong Li, Yuankai Qi, Ziheng Yan, Zhenjun Han, Anton van den Hengel, Ming-Hsuan Yang, Qingming Huang
Researchers introduce Weakly Supervised Video Individual Counting (WVIC), a framework that accurately determines the total count of unique individuals in a video using only sparse inflow/outflow annotations. The proposed CGNet model achieves superior performance, improving Mean Absolute Error by up to 46% on the CroHD dataset compared to fully supervised methods, while reducing annotation costs by 70%.
Key insights
- High accuracy in video individual counting can be achieved with significantly weaker supervision (inflow/outflow labels), indicating that costly trajectory labels are not strictly necessary.
- The Group-level Matching Loss (GML), based on Optimal Transport, effectively encourages discriminative feature learning by pushing features of shared individuals closer and unshared individuals apart.
- A Memory-based Individual Count Predictor (MCP) improves long-term counting by dynamically updating and pruning individual templates, robustly handling appearance changes and re-entering individuals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.