Research map / Architectures and efficiency / Neural architecture search
Searching for Two-Stream Models in Multivariate Space for Video Recognition
Xinyu Gong, Heng Wang, Mike Zheng Shou, Matt Feiszli, Zhangyang Wang, Zhicheng Yan
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 |
|---|---|---|---|
| add a fast stream and searched fusion blocks to the slow-stream model (ablation) | Auto-TSNet · Kinetics-400 | top-1 accuracy (%): 70.8 -> 74.1 (+3.3) | Improved |
| add searched GloRe attention blocks (ablation) | Auto-TSNet step-2 model · Kinetics-400 | top-1 accuracy (%): 74.1 -> 74.6 (+0.5) | Improved |
| search the backbone instead of hand-crafting it (ablation) | Slow stream · Kinetics-400 | top-1 accuracy (%): 70.2 -> 70.8 | Improved |
| search fusion locations instead of manual placement (ablation) | Manual-TSNet-70% · Kinetics-400 | top-1 accuracy (%): 73.2 -> 73.7 | Improved |
| search GloRe locations instead of no attention (ablation) | Manual-TSNet-70% · Kinetics-400 | top-1 accuracy (%): 73.2 -> 73.6 | Improved |
| use progressive search instead of one-step search (ablation) | Auto-TSNet-S · Kinetics-400 | top-1 accuracy (%): 72.8 -> 75.4 | Improved |
| use progressive search instead of one-step search (search time) (ablation) | Auto-TSNet-S · Kinetics-400 | search days: 6.5 -> 4.8 | Improved |
| replace single-stream X3D-S with a manual two-stream model (70% FLOPS to sparse stream) | Manual-TSNet · Kinetics-400 | top-1 accuracy (%): 72.9 -> 73.2 | Improved |
| lower the sparse-stream FLOPS ratio from 70% to 55% | Manual-TSNet · Kinetics-400 | top-1 accuracy (%): 73.2 -> 72.4 | Hurt |
| increase manual fusion blocks from 5 to 10 (ablation) | Manual-TSNet-70% · Kinetics-400 | top-1 accuracy (%): 73.2 -> 71.2 | Hurt |
| add 5 manual fusion blocks to a model with none (ablation) | Manual-TSNet-70% · Kinetics-400 | top-1 accuracy (%): 70.7 -> 73.2 | Improved |
| manually add two GloRe blocks in stage 2 only (ablation) | Manual-TSNet-70% · Kinetics-400 | top-1 accuracy (%): 73.2 -> 72.7 | Hurt |
Closest papers by meaning
- AssembleNet: Searching for Multi-Stream Neural Connectivity in Video Architectures (ICLR 2020 · similarity 0.88)
- Convolutional Two-Stream Network Fusion for Video Action Recognition (CVPR 2016 · similarity 0.87)
- BossNAS: Exploring Hybrid CNN-Transformers With Block-Wisely Self-Supervised Neural Architecture Search (ICCV 2021 · similarity 0.86)
- MVFNet: Multi-View Fusion Network for Efficient Video Recognition (AAAI 2021 · similarity 0.85)
- Semantic Adversarial Network with Multi-Scale Pyramid Attention for Video Classification (AAAI 2019 · similarity 0.85)
- BossNAS Family: Block-Wisely Self-Supervised Neural Architecture Search (TPAMI 2025 · similarity 0.85)
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.
