Reproduce “Counting Beyond Instances: A Benchmark for Group-Individual Object 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
Rui Wang, Junyi Huang, Jiahui Li, Qiao Yu, Yixue Hao, Long Hu, Baoru Huang
This research introduces Group-Individual Object Counting (GIC), a paradigm where models count objects at both individual and semantic group levels based on a specified unit. Researchers developed the BunchCount dataset with paired annotations and a framework called GICount, which reduces group-level counting errors by 48.5% compared to prior baselines while maintaining strong individual-level performance.
Key insights
- Existing object counting models exhibit a "semantic granularity conflict," where optimizing for individual counting often degrades group counting performance and vice versa, indicating distinct challenges for each granularity.
- Explicitly modeling group-individual containment relations during training is crucial for achieving accurate semantic-group counting.
- Disentangling category semantics from group-structure semantics within the counting-unit prompt is essential to prevent cross-granularity interference and maintain balanced performance across individual and group counting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.