Research map / Video / Action recognition
AdaScan: Adaptive Scan Pooling in Deep Convolutional Neural Networks for Human Action Recognition in Videos
Check 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 |
|---|---|---|---|
| Replace mean pooling with adaptive importance-weighted temporal pooling | VGG-16 spatial network · UCF101, average over 3 splits | classification accuracy: 77.6 to 78.6 | Helped |
| Replace mean pooling with adaptive importance-weighted temporal pooling | VGG-16 temporal network · UCF101, average over 3 splits | classification accuracy: 82.4 to 83.4 | Helped |
| Replace mean pooling with adaptive importance-weighted temporal pooling | VGG-16 spatial network · HMDB51, average over 3 splits | classification accuracy: 40.9 to 41.4 | Helped |
| Replace mean pooling with adaptive importance-weighted temporal pooling | VGG-16 temporal network · HMDB51, average over 3 splits | classification accuracy: 48.6 to 49.2 | Helped |
| Use a residual difference between current-frame and pooled features as the importance-network input (ablation) | AdaScan · HMDB51 and UCF101 | classification performance | Helped |
| Replace the entropy-based importance regularizer with an ℓ1 regularizer (ablation) | AdaScan · video action classification | classification performance | Hurt |
| Initialize the pooled state from the first-frame features rather than a random vector (ablation) | AdaScan · video action classification | training stability | Helped |
| Randomly reinitialize the Adaptive Pooling module instead of transferring its weights from UCF101 (ablation) | AdaScan trained on HMDB51 · HMDB51 | classification performance | Helped |
| Increase entropy-regularization strength from 1e6 to 1e7 (ablation) | AdaScan spatial network · HMDB51 | frame-selection sparsity | Helped |
| Increase entropy-regularization strength beyond 1e7 (ablation) | AdaScan spatial network · HMDB51 | classification performance | Hurt |
| Fuse AdaScan with improved dense trajectories | AdaScan · UCF101 | classification accuracy: 89.4 to 91.3 | Helped |
| Fuse AdaScan with improved dense trajectories | AdaScan · HMDB51 | classification accuracy: 54.9 to 61.0 | Helped |
Closest papers by meaning
- Video Representation Learning Using Discriminative Pooling (CVPR 2018 · similarity 0.93)
- Dynamic Image Networks for Action Recognition (CVPR 2016 · similarity 0.89)
- Learning End-to-end Video Classification with Rank-Pooling (ICML 2016 · similarity 0.89)
- Attentional Pooling for Action Recognition (NeurIPS 2017 · similarity 0.88)
- More Is Less: Learning Efficient Video Representations by Big-Little Network and Depthwise Temporal Aggregation (NeurIPS 2019 · similarity 0.88)
- Weakly Supervised Action Localization by Sparse Temporal Pooling Network (CVPR 2018 · 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.