Research map / Detection and segmentation / CNN features
Rethinking Spatial Invariance of Convolutional Networks for Object Counting
Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, Alexander G. Hauptmann
Check 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 |
|---|---|---|---|
| replace convolution filters with low-rank Gaussian convolutions in ResNet-50 | ResNet-50 · SHTech-PartA | MAE: 55.4 -> 54.8 | Improved |
| replace convolution filters with low-rank Gaussian convolutions in ResNet-50 | ResNet-50 · UCF_CC_50 | MAE: 197.5 -> 186.3 | Improved |
| replace convolution filters with low-rank Gaussian convolutions in CSRNet | CSRNet · UCF-QNRF | MAE: 71.3 -> 84.2 | Hurt |
| replace convolution filters with low-rank Gaussian convolutions in ResNet-50 | ResNet-50 · JHU-CROWD++ | MAE: 59.9 -> 58.2 | Improved |
| replace convolution filters with low-rank Gaussian convolutions in ResNet-50 | ResNet-50 · SHTech-PartB | MAE: 6.4 -> 6.2 | Improved |
| replace convolution filters with low-rank Gaussian convolutions in ResNet-50 | ResNet-50 · TRANCOS | MAE: 2.6 -> 2.1 | Improved |
| replace convolution filters with low-rank Gaussian convolutions in ResNet-50 | ResNet-50 · MTC | MAE: 5.6 -> 3.1 | Improved |
| add low-rank approximation to Gaussian convolution (ablation) | Gaussian convolution on MCNN · MCNN, batchsize 1 and image size 256² | forward time (milliseconds): 51.3 -> 13.3 | Improved |
| add translation invariance to low-rank Gaussian convolution (ablation) | LRA on MCNN · MCNN, batchsize 1 and image size 256² | forward time (milliseconds): 13.3 -> 4.1 | Improved |
| add low-rank approximation to Gaussian convolution (ablation) | Gaussian convolution on MCNN · MCNN, batchsize 1 and image size 256² | backward time (milliseconds): 160.0 -> 44.1 | Improved |
| add translation invariance to low-rank Gaussian convolution (ablation) | LRA on MCNN · MCNN, batchsize 1 and image size 256² | backward time (milliseconds): 44.1 -> 12.6 | Improved |
| use larger K in spatial pyramid pooling (ablation) | GauNet in MCNN and SANet · MCNN and SANet spatial pyramid pooling | counting performance | Mixed |
Closest papers by meaning
- Counting With Focus for Free (ICCV 2019 · similarity 0.90)
- Focus for Free in Density-Based Counting (IJCV 2024 · similarity 0.89)
- From Open Set to Closed Set: Supervised Spatial Divide-and-Conquer for Object Counting (IJCV 2023 · similarity 0.89)
- Learning Spatial Awareness to Improve Crowd Counting (ICCV 2019 · similarity 0.89)
- AutoScale: Learning to Scale for Crowd Counting (IJCV 2022 · similarity 0.88)
- From Open Set to Closed Set: Counting Objects by Spatial Divide-and-Conquer (ICCV 2019 · 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.
