Research map / Video / Action recognition
Quantifying and Learning Static vs. Dynamic Information in Deep Spatiotemporal Networks
Matthew Kowal, Mennatullah Siam, Md Amirul Islam, Neil D. B. Bruce, Richard P. Wildes, Konstantinos G. Derpanis
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 fast branch to the slow pathway (SlowFast vs SlowOnly) | SlowFast · Kinetics-400 action recognition, stage 2 | dynamic units in slow pathway: +3.3% | Improved |
| train on SSv2 instead of Kinetics-400 or Diving48 (dataset choice) | SlowOnly and FastOnly · Action recognition datasets | dynamic units | Improved |
| train with shuffled frames (static-only learning) (ablation) | SlowOnly, FastOnly · SSv2 vs Diving48 vs Kinetics-400 | relative Top-1 accuracy | Mixed |
| train on TAO-VOS instead of DAVIS, YouTube-VOS or ImageNetVID | FusionSeg · AVOS fusion layer 5 | static bias (%): 30.1 -> 25.8 (DAVIS vs TAO-VOS) | Improved |
| train on OVIS instead of YTVIS | VisTR-R50 · VIS encoder layer six | dynamic bias | Improved |
| remove top-k dynamic units vs static units at inference (ablation) | SlowFast · SSv2 top-1 accuracy | accuracy drop: 70% of dynamic units removed: 5% more drop than static | Hurt |
| remove 40% of most static units in fusion layer 4 (ablation) | MATNet · MoCA | mIoU: 0% when removing only 40% | Hurt |
| add StaticDropout (rates 0.1-0.7) vs standard or no dropout | 3D-ResNet50 / 18 · SSv2 and Diving48 | ratio of dynamic units | Improved |
| add StaticDropout vs baselines | 3D-ResNet50 · SSv2 and SomethingElse | Top-1 accuracy | Improved |
| raise StaticDropout rate to r=0.7 (ablation) | StaticDropout · SSv2 and SomethingElse | Top-1 accuracy | Hurt |
| StaticDropout (r=0.3) vs standard dropout on rare vs common classes | 3D-ResNet50 · SomethingElse compositional split | per-class accuracy | Mixed |
| use bidirectional (B) instead of unidirectional (U) cross connections with gated fusion (ablation) | MATNet/RTNet variants · AVOS fusion layers 4 and 5 | dynamic bias | Hurt |
Closest papers by meaning
- A Deeper Dive Into What Deep Spatiotemporal Networks Encode: Quantifying Static vs. Dynamic Information (CVPR 2022 · similarity 0.97)
- Mitigating and Evaluating Static Bias of Action Representations in the Background and the Foreground (ICCV 2023 · similarity 0.90)
- What Have We Learned From Deep Representations for Action Recognition? (CVPR 2018 · similarity 0.88)
- Deep Insights into Convolutional Networks for Video Recognition (IJCV 2019 · similarity 0.88)
- A Large-scale Study of Spatiotemporal Representation Learning with a New Benchmark on Action Recognition (ICCV 2023 · similarity 0.88)
- Bias for Action: Video Implicit Neural Representations with Bias Modulation (CVPR 2025 · similarity 0.86)
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.
