Research map / Video / Action recognition
Spatiotemporal Residual Networks for Video Action Recognition
Read the paperCheck 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 |
|---|---|---|---|
| add cross-stream residual connections and temporal convolutions (ST-ResNet) to two-stream ResNet-50 (ablation) | two-stream ResNet-50 · UCF101 | mean accuracy: 89.47% -> 92.76% | Helped |
| add cross-stream residual connections and temporal convolutions (ST-ResNet) to two-stream ResNet-50 (ablation) | two-stream ResNet-50 · HMDB51 | mean accuracy: 60.59% -> 65.57% | Helped |
| add temporal max-pooling over 5 chunks after pool5 (ST-ResNet*) (ablation) | ST-ResNet · UCF101 | mean accuracy: 92.76% -> 93.46% | Helped |
| add temporal max-pooling over 5 chunks after pool5 (ST-ResNet*) (ablation) | ST-ResNet · HMDB51 | mean accuracy: 65.57% -> 66.41% | Helped |
| use temporal max-pooling instead of temporal average pooling after pool5 (ablation) | ST-ResNet* · UCF101 and HMDB51 | mean accuracy | Helped |
| reduce batch-norm sample count for appearance stream (batch size 4 for moment estimation) (ablation) | appearance stream · UCF101 | validation accuracy: approximately 4% gain | Helped |
| use dropout after classification layer in appearance stream (ablation) | appearance stream · UCF101 | validation accuracy | Hurt |
| use a single joint softmax classifier instead of separate loss layers in ST-ResNet (ablation) | ST-ResNet · UCF101 | validation accuracy | Hurt |
| add direct/bidirectional cross-stream connections between identical layers instead of the chosen additive residual from motion to appearance (ablation) | two-stream architecture · preliminary experiments | validation error | Hurt |
| average non-softmaxed prediction scores for stream fusion instead of softmax scores (ablation) | two-stream fusion · UCF101 and HMDB51 | accuracy | Helped |
| add IDT features fused with ST-ResNet* predictions (ablation) | ST-ResNet* · HMDB51 | mean accuracy: 66.4% -> 70.3% | Helped |
| add IDT features fused with ST-ResNet* predictions (ablation) | ST-ResNet* · UCF101 | mean accuracy: 93.4% -> 94.6% | Helped |
Closest papers by meaning
- Convolutional Two-Stream Network Fusion for Video Action Recognition (CVPR 2016 · similarity 0.93)
- Spatiotemporal Pyramid Network for Video Action Recognition (CVPR 2017 · similarity 0.92)
- A Closer Look at Spatiotemporal Convolutions for Action Recognition (CVPR 2018 · similarity 0.92)
- Spatiotemporal Multiplier Networks for Video Action Recognition (CVPR 2017 · similarity 0.90)
- StNet: Local and Global Spatial-Temporal Modeling for Action Recognition (AAAI 2019 · similarity 0.90)
- Temporal Residual Networks for Dynamic Scene Recognition (CVPR 2017 · 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.