Research map / Label-efficient and robust learning / Recognition and long-tail
NWPU-Crowd: A Large-Scale Benchmark for Crowd Counting and Localization
Qi Wang, Junyu Gao, Wei Lin, Xuelong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with denser positive samples alongside negatives (S0+S1 to S0+S3+S4) (ablation) | C3F-VGG · NWPU-Crowd (val), S0 MAE | MAE: 18.54 -> 147.53 | Hurt |
| train each density class separately instead of together (ablation) | C3F-VGG · NWPU-Crowd (val), S1 | MAE: 33.12 -> 21.00 | Improved |
| train each density class separately instead of together (ablation) | C3F-VGG · NWPU-Crowd (val), S3+S4 (S4 column) | MAE: 3997.57 -> 3488.25 | Improved |
| add negative samples (S0) to S1 training (ablation) | C3F-VGG · NWPU-Crowd (val), S1 | MAE: 21.00 -> 16.44 | Improved |
| add negative samples (S0) to S2 training (ablation) | C3F-VGG · NWPU-Crowd (val), S2 | MAE: 51.01 -> 49.97 | Improved |
| add negative samples (S0) to S3+S4 training (ablation) | C3F-VGG · NWPU-Crowd (val), S4 | MAE: 3488.25 -> 2882.23 | Improved |
| add negative samples (S0) to S1+...+S4 training (ablation) | C3F-VGG · NWPU-Crowd (val), S4 | MAE: 3997.57 -> 3730.56 | Improved |
| use 100% instead of 10% of training data (ablation) | C3F-VGG · NWPU-Crowd (val) | MAE: 158.35 -> 105.79 | Improved |
| evaluate on test set instead of val set | SCAR · NWPU-Crowd | MAE: 81.57 -> 110.0 | Hurt |
| use a stricter localization threshold (sigma_s instead of sigma_l) | RAZ_Loc · NWPU-Crowd (val) localization | F1-measure: 62.5 -> 54.5 | Hurt |
| use box-supervised detector (Faster RCNN) instead of point-supervised RAZ_Loc | localization · NWPU-Crowd (test) | F1-measure: 59.8 -> 6.7 | Hurt |
Closest papers by meaning
- Composition Loss for Counting, Density Map Estimation and Localization in Dense Crowds (ECCV 2018 · similarity 0.90)
- Rethinking Counting and Localization in Crowds: A Purely Point-Based Framework (ICCV 2021 · similarity 0.90)
- SynMVCrowd: A Large Synthetic Benchmark for Multi-view Crowd Counting and Localization (IJCV 2026 · similarity 0.89)
- AutoScale: Learning to Scale for Crowd Counting (IJCV 2022 · similarity 0.89)
- JHU-CROWD++: Large-Scale Crowd Counting Dataset and A Benchmark Method (TPAMI 2020 · similarity 0.89)
- Locate, Size and Count: Accurately Resolving People in Dense Crowds via Detection (TPAMI 2020 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
