Research map / Video / Action recognition
Convolutional Two-Stream Network Fusion for Video Action Recognition
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 |
|---|---|---|---|
| Use 3D Conv fusion and 3D pooling across streams | Two-stream VGG action recognition · UCF101 (split 1) | accuracy: 89.64% -> 90.40% | Helped |
| Use 3D Conv fusion and 3D pooling across streams | Two-stream VGG action recognition · HMDB51 (split 1) | accuracy: 57.58% -> 58.63% | Helped |
| Replace 2D pooling with 3D pooling (ablation) | Two-stream VGG action recognition with 2D Conv fusion · UCF101 (split 1) | accuracy: 89.35% -> 89.64% | Helped |
| Replace 2D pooling with 3D pooling (ablation) | Two-stream VGG action recognition with 2D Conv fusion · HMDB51 (split 1) | accuracy: 56.93% -> 57.58% | Helped |
| Keep both streams after fusion instead of truncating one (ablation) | Two-stream VGG-16 fusion · UCF101 | mean classification accuracy: 91.8% -> 92.5% | Helped |
| Keep both streams after fusion instead of truncating one (ablation) | Two-stream VGG-16 fusion · HMDB51 | mean classification accuracy: 64.6% -> 65.4% | Helped |
| Fuse at ReLU5 rather than earlier convolution layers (ablation) | Two VGG-M-2048 streams · UCF101 (split 1) | accuracy: 82.25% -> 85.96%; 83.43% -> 85.96%; 82.55% -> 85.96% | Helped |
| Add fusion at FC8 alongside ReLU5 fusion (ablation) | Two VGG-M-2048 streams · UCF101 (split 1) | accuracy: 85.96% -> 86.04% | Helped |
| Add early fusion at ReLU3 alongside ReLU5 and FC6 (ablation) | Two VGG-M-2048 streams · UCF101 (split 1) | accuracy: 85.96% -> 81.55% | Hurt |
| Use Conv fusion rather than sum fusion at ReLU5 (ablation) | Two VGG-M-2048 streams · UCF101 (split 1) | accuracy: 85.20% -> 85.96% | Helped |
| Use Conv fusion rather than max or concatenation fusion at ReLU5 (ablation) | Two VGG-M-2048 streams · UCF101 (split 1) | accuracy: 82.70% -> 85.96%; 83.53% -> 85.96% | Helped |
| Fuse at ReLU5 rather than at the softmax layer (ablation) | Two VGG-M-2048 streams · UCF101 (split 1) | accuracy: 85.94% -> 85.96% | No clear effect |
Closest papers by meaning
- Spatiotemporal Residual Networks for Video Action Recognition (NeurIPS 2016 · similarity 0.93)
- Spatiotemporal Pyramid Network for Video Action Recognition (CVPR 2017 · similarity 0.92)
- MVFNet: Multi-View Fusion Network for Efficient Video Recognition (AAAI 2021 · similarity 0.90)
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.90)
- Spatiotemporal Multiplier Networks for Video Action Recognition (CVPR 2017 · similarity 0.90)
- Video Modeling With Correlation Networks (CVPR 2020 · similarity 0.89)
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.