Research map / Label-efficient and robust learning / Recognition and long-tail
To Choose or to Fuse? Scale Selection for Crowd Counting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use patch-wise adaptive feature-level selection with confidence-weighted averaging instead of averaging all level predictions (ablation) | SASNet · ShanghaiTech PartA | MAE: 57.48 -> 54.75 | Helped |
| Select the feature level with minimum counting error using ground truth (ablation) | SASNet · ShanghaiTech PartA | MAE: 57.48 -> 46.19 | Helped |
| Use only the highest-confidence feature level rather than confidence-weighted averaging (ablation) | SASNet · ShanghaiTech PartA | MAE: 54.75 -> 55.71 | Hurt |
| Add PRA Loss to a top-feature-level baseline (ablation) | SASNet baseline · ShanghaiTech PartA | MAE: 61.04 -> 57.50 | Helped |
| Add PRA Loss to adaptive feature-level selection (ablation) | SASNet · ShanghaiTech PartA | MAE: 54.75 -> 53.59 | Helped |
| Use a 32 × 32 rather than 8 × 8 or 16 × 16 selection patch (ablation) | SASNet · ShanghaiTech PartA | MAE: 55.55 -> 54.75; 55.38 -> 54.75 | Helped |
| Use a 64 × 64 rather than 32 × 32 selection patch (ablation) | SASNet · ShanghaiTech PartA | MAE: 54.75 -> 55.54 | Hurt |
| Use SASNet rather than the strongest listed baseline | crowd counting methods · ShanghaiTech PartA | MAE: 55.4 -> 53.59 | Helped |
| Use SASNet rather than the strongest listed baseline | crowd counting methods · ShanghaiTech PartB | MAE: 6.5 -> 6.35 | Helped |
| Use SASNet rather than the strongest listed baseline | crowd counting methods · UCF CC 50 | MAE: 174.84 -> 161.4 | Helped |
| Use SASNet rather than the strongest listed baseline | crowd counting methods · UCF-QNRF | MAE: 71.3 -> 85.2 | Hurt |
| Use SASNet rather than the strongest listed baseline | crowd counting methods · WorldExpo’10 | average MAE: 6.64 -> 5.71 | Helped |
Closest papers by meaning
- Counting without Scale: Scale-Consistent Error Correction for Crowd Counting (NeurIPS 2026 · similarity 0.90)
- Crowd Counting via Adversarial Cross-Scale Consistency Pursuit (CVPR 2018 · similarity 0.88)
- Attention Scaling for Crowd Counting (CVPR 2020 · similarity 0.88)
- Shallow Feature Based Dense Attention Network for Crowd Counting (AAAI 2020 · similarity 0.88)
- Free Lunch Enhancements for Multi-modal Crowd Counting (CVPR 2025 · similarity 0.86)
- Taste More, Taste Better: Diverse Data and Strong Model Boost Semi-Supervised Crowd Counting (CVPR 2025 · similarity 0.85)
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.