Research map / Label-efficient and robust learning / Self-supervised and contrastive
Completely Self-Supervised Crowd Counting via Distribution Matching
Deepak Babu Sam, Abhinav Agarwalla, Jimmy Joseph, Vishwanath A. Sindagi, R. Venkatesh Babu, Vishal M. Patel
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 |
|---|---|---|---|
| add Stage 1 rotation self-supervised pretraining before Sinkhorn training (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 257.5 -> 197.3 | Improved |
| use rotation pretext instead of colorization (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 242.5 -> 197.3 | Improved |
| use rotation self-supervision instead of ImageNet weights for the FEN (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 257.8 -> 197.3 | Improved |
| use VGG-style FEN instead of ResNet blocks (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 244.8 -> 197.3 | Improved |
| use truncated power-law prior instead of uniform prior (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 261.8 -> 197.3 | Improved |
| add skip features from the second conv block to the regressor (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 226.8 -> 197.3 | Improved |
| use 4 rotation classes instead of 2 (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 233.5 -> 197.3 | Improved |
| use 8x8 cells instead of 4x4 for count histograms (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 251.6 -> 197.3 | Improved |
| use ground-truth count distribution instead of the parametric prior (ablation) | CSS-CCNN · ShanghaiTech Part A (single run) | MAE: 197.3 vs 202.7 | No clear effect |
| add edge-based pseudo density grouping to Sinkhorn loss (CSS-CCNN++) | CSS-CCNN · ShanghaiTech Part A | MAE: 207.3 -> 195.6 | Improved |
| add edge-based pseudo density grouping to Sinkhorn loss (CSS-CCNN++) | CSS-CCNN · UCF-QNRF | MAE: 442.4 -> 414.0 | Improved |
| add edge-based pseudo density grouping to Sinkhorn loss (CSS-CCNN++) | CSS-CCNN · JHU-CROWD++ | MAE: 243.6 -> 197.9 | Improved |
Closest papers by meaning
- Active Crowd Counting with Limited Supervision (ECCV 2020 · similarity 0.92)
- Semi-Supervised Crowd Counting via Self-Training on Surrogate Tasks (ECCV 2020 · similarity 0.92)
- Leveraging Self-Supervision for Cross-Domain Crowd Counting (CVPR 2022 · similarity 0.91)
- Almost Unsupervised Learning for Dense Crowd Counting (AAAI 2019 · similarity 0.90)
- Semi-Supervised Counting via Pixel-by-Pixel Density Distribution Modeling (TPAMI 2025 · similarity 0.90)
- CrowdDiff: Multi-hypothesis Crowd Density Estimation using Diffusion Models (CVPR 2024 · similarity 0.90)
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.
