Research map / Video / Action recognition
Action Recognition with Dynamic Image Networks
Hakan Bilen, Basura Fernando, Efstratios Gavves, Andrea Vedaldi
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 |
|---|---|---|---|
| replace mean or max pooling with rank pooling for the single-image video summary (ablation) | single dynamic image · UCF101 split-1 | accuracy (%): 52.6 -> 57.2 (mean), 48.0 -> 57.2 (max) | Improved |
| use dynamic image instead of motion history image (ablation) | single image CaffeNet · UCF101 split-1 | accuracy (%): 46.6 -> 57.2 | Improved |
| use approximate rank pooling instead of exact rank pooling | dynamic image · UCF101 split-1 | fps: 131 -> 5920 | Improved |
| use approximate rank pooling instead of exact rank pooling | dynamic image · UCF101 split-1 | classification accuracy (%): 57.9 -> 55.2 | Hurt |
| use approximate rank pooling instead of exact rank pooling | dynamic image · UCF101 split-1 | pairwise ranking accuracy (%): 95.5 -> 96.5 | No clear effect |
| split video into multiple dynamic images (window 10, stride 6) instead of one per video (ablation) | MDI · UCF101 split-1 | accuracy (%): 55.2 -> 67.4 | Improved |
| increase window size to 20 with stride 12 (ablation) | MDI · UCF101 split-1 | accuracy (%): 67.4 -> 58.9 | Hurt |
| use max pooling over dynamic images instead of mean or ARP in temporal pooling layer (ablation) | MDI · UCF101 split-1 | accuracy (%): 66.2 -> 68.3, 65.2 -> 68.3 | Improved |
| move ARP from input (conv1) to before conv3 (ablation) | dynamic map network · UCF101 | accuracy (%): 70.9 -> 73.3 | Improved |
| place ARP before conv4 instead of conv3 (ablation) | dynamic map network · UCF101 | accuracy (%): 73.3 -> 67.6 | Hurt |
| learn pooling coefficients (parametric pooling) instead of fixed ARP (ablation) | dynamic image network · UCF101 and HMDB51 | accuracy (%): ARP conv1 70.9 vs PP1 67.6; ARP conv2 68.5 vs PP2 73.9 | No clear effect |
| switch CaffeNet to ResNeXt-50 | SI and DI streams · UCF101 | accuracy (%): 70.9 -> 86.6 (DI) | Improved |
Closest papers by meaning
- Dynamic Image Networks for Action Recognition (CVPR 2016 · similarity 0.99)
- Discriminatively Learned Hierarchical Rank Pooling Networks (IJCV 2017 · similarity 0.93)
- Rank Pooling for Action Recognition (TPAMI 2016 · similarity 0.91)
- Discriminative Hierarchical Rank Pooling for Activity Recognition (CVPR 2016 · similarity 0.91)
- Second-order Temporal Pooling for Action Recognition (IJCV 2018 · similarity 0.90)
- Improving Video Model Transfer With Dynamic Representation Learning (CVPR 2022 · 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.
