Research map / Architectures and efficiency / CNN design
Perspective-Guided Convolution Networks for Crowd Counting
Zhaoyi Yan, Yuchen Yuan, Wangmeng Zuo, Xiao Tan, Yezhen Wang, Shilei Wen, Errui Ding
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 perspective-guided convolution blocks (5) to CSRNet* baseline | CSRNet* · ShanghaiTech Part A | MAE: 67.5 -> 57.0 | Improved |
| add 3 PGC blocks to ResNet-101 backbone (ablation) | ResNet-101 · ShanghaiTech Part A / B | MAE: 109.6/26.2 -> 89.7/18.6 | Improved |
| increase Gaussian filter size K from 1 to 7 (ablation) | single PGC block · ShanghaiTech Part A | MAE: 67.3 -> 65.8 | Improved |
| increase Gaussian filter size K from 7 to 9 or larger (ablation) | single PGC block · ShanghaiTech Part A | MAE: 65.8 -> 66.3 (K=9), 66.4 (K=11) | Hurt |
| stack 5 PGC blocks instead of 1 (ablation) | PGCNet · ShanghaiTech Part A / B | MAE: 65.8/9.8 -> 57.0/8.8 | Improved |
| stack 6 PGC blocks instead of 5 (ablation) | PGCNet · ShanghaiTech Part A / B | MAE: 57.0/8.8 -> 58.3/9.0 | Hurt |
| pre-train PENet (ablation) | PGCNet · UCF_CC_50 | MAE: 278.6 -> 244.6 | Improved |
| pre-train PENet (ablation) | PGCNet · Crowd Surveillance | MAE: 10.3 -> 7.2 | Improved |
| train PENet end-to-end (Ours B) instead of using fixed estimated perspective (Ours A) (ablation) | PGCNet · UCF_CC_50 | MAE: 259.4 -> 244.6 | Improved |
| train PENet end-to-end (Ours B) instead of using fixed estimated perspective (Ours A) (ablation) | PGCNet · UCF_CC_50 | MSE: 317.6 -> 361.2 | Hurt |
| train PENet end-to-end (Ours B) instead of Ours A (ablation) | PGCNet · Crowd Surveillance | MAE: 7.7 -> 7.2 | Improved |
| use ground-truth instead of PENet-estimated perspective map (ablation) | PGCNet · ShanghaiTech Part A / B, WorldExpo'10 | MAE: 58.1/9.0/8.3 -> 57.0/8.8/8.1 | Improved |
Closest papers by meaning
- Revisiting Perspective Information for Efficient Crowd Counting (CVPR 2019 · similarity 0.93)
- Single-Image Crowd Counting via Multi-Column Convolutional Neural Network (CVPR 2016 · similarity 0.91)
- Leveraging Self-Supervision for Cross-Domain Crowd Counting (CVPR 2022 · similarity 0.89)
- Context-Aware Crowd Counting (CVPR 2019 · similarity 0.89)
- Attention Scaling for Crowd Counting (CVPR 2020 · similarity 0.88)
- Incorporating Side Information by Adaptive Convolution (NeurIPS 2017 · similarity 0.88)
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.
