Research map / Label-efficient and robust learning / Semi-supervised learning
Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes point-to-region (P2R) matching, replacing P2P in a pseudo-labeling framework for point-based crowd counting, so confidence propagates to background pixels; supports semi-supervised counting and UDA.
Problem
Point-based crowd counters need many point annotations per image; semi-supervised pseudo-labeling with P2P matching breaks down because confidence cannot be assigned to background pixels, causing training to fail.
Approach
Introduce a point-specific activation map (PSAM) to show that unlabeled losses over-activate neighborhoods of foreground pixels, then replace Hungarian-based P2P matching with P2R, which matches local regions to pseudo-points so pixels in each region share the pseudo-point's confidence and background supervision is retained.
Key results
- With 5% labels on ShTech A, P2R reaches 69.9 MAE versus 83.7 for OT-M and 85.4 for DAC.
- Fully supervised P2R improves over P2PNet on all four datasets (e.g., ShTech A MAE 51.02 vs 52.74).
- P2R loss computation takes 0.0064 s versus 0.4307 s for P2P on a 775-point image, about 68x faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace P2P matching with P2R matching | Point-based semi-supervised counter · ShTech A, ShTech B, UCF-QNRF, and JHU++ | MAE | Helped |
| Replace P2P matching with P2R matching | Fully supervised point-based counter · ShTech A, ShTech B, UCF-QNRF, and JHU++ | MAE | Helped |
| Add target-domain unlabeled data adaptation (ablation) | P2R UDA · A→B, A→Q, B→A, and B→Q | MAE | Helped |
| Use P2R instead of P2P matching (ablation) | Semi-supervised point-based counter · ShTech A 5% labeled | MAE | Helped |
| Vary τ for selecting the foreground pixel in a matched region (ablation) | P2R · ShTech A fully supervised | MAE | Mixed |
| Set τ to 8 (ablation) | P2R · ShTech A fully supervised | MAE | Helped |
| Vary confidence threshold η (ablation) | P2R · ShTech A 5% labeled | estimation error | Mixed |
| Set confidence threshold η to 0.7 (ablation) | P2R · ShTech A 5% labeled | estimation error | Helped |
| Increase point-region radius μ from 32 to 256 (ablation) | P2R · ShTech A 5% labeled | MAE | No clear effect |
| Vary the unlabeled-loss weight α (ablation) | P2R · ShTech A 5% labeled | MAE | Mixed |
| Set α/(1−α) to 2 (ablation) | P2R · ShTech A 5% labeled | MAE | Helped |
| Use unlabeled data in training (α greater than 0) (ablation) | P2R · ShTech A 5% labeled | MAE | Helped |
Surprising findings
- Under P2P pseudo-labeling, training with unlabeled data causes failure: validation MAE/MSE increase once unlabeled data is introduced.
- Supervised P2R with 100% labels also outperforms P2PNet across all datasets.
Limitations
- Semi-supervised threshold and radius hyperparameters (eta, mu, tau, alpha) need tuning per setting.
- Failure case due to crowd motion blur is noted.
Experimental setup
Datasets: ShanghaiTech A ShanghaiTech B UCF-QNRF JHU++
Benchmarks: ShanghaiTech A ShanghaiTech B UCF-QNRF JHU++
Metrics: MAE MSE
Baselines: label-only MT L2R GP DAC OT-M IRAST SUA CU P2PNet DCCUS MPCount
Builds on: P2PNet mean teacher FixMatch Grad-CAM OT-M
Closest papers by meaning
- Optimal Transport Minimization: Crowd Localization on Density Maps for Semi-Supervised Counting (CVPR 2023 · similarity 0.89)
- Recurrent Attentive Zooming for Joint Crowd Counting and Precise Localization (CVPR 2019 · similarity 0.88)
- Counting without Scale: Scale-Consistent Error Correction for Crowd Counting (NeurIPS 2026 · similarity 0.87)
- Learning to Count via Unbalanced Optimal Transport (AAAI 2021 · similarity 0.87)
- Point in, Box Out: Beyond Counting Persons in Crowds (CVPR 2019 · similarity 0.87)
- Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting (CVPR 2025 · similarity 0.84)
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.