Research map / Video / Action recognition
Eventful Transformers: Leveraging Temporal Redundancy in Vision Transformers
Matthew Dutson, Yin Li, Mohit Gupta
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 temporal redundancy awareness (token gating, r=768) to ViTDet | ViTDet on VID, 1024 · ImageNet VID val | mAP50 / GFlops: 82.93 -> 81.25 mAP50; 467.4 -> 122.3 GFlops | Improved |
| use reference-based gating instead of STGT lossy gating (ablation) | Eventful ViTDet · ImageNet VID val (1024, r=256) | mAP50: 68.73 -> 73.29 | Improved |
| extend redundancy awareness to query-key and attention-value products (ablation) | Eventful ViTDet · ImageNet VID val (1024, r=256) | GFlops: 142.3 -> 53.3 | Improved |
| extend redundancy awareness to attention products at low r (accuracy side) (ablation) | Eventful ViTDet · ImageNet VID val (1024, r=256) | mAP50: 76.60 -> 73.29 | Hurt |
| use threshold policy instead of top-r policy (ablation) | Eventful ViTDet · ImageNet VID val (1024) | mAP50 vs GFlops: h=5.0: 78.11 at 133.5 GFlops vs r=512: 79.38 at 87.8 GFlops | Hurt |
| add spatial pooling of k and v to the temporal model (ablation) | Eventful ViTDet · ImageNet VID val (1024) | GFlops: 388.1 -> 217.0 GFlops at r=2048 (80.15 -> 80.14 mAP50) | Improved |
| naively replace ViViT spatial model with Eventful version without fine-tuning temporal model (ablation) | ViViT on EPIC-Kitchens · EPIC-Kitchens 100 | accuracy: about -10% at r=100 | Hurt |
| fine-tune non-Eventful temporal model on outputs of frozen Eventful spatial model (ablation) | Eventful ViViT · EPIC-Kitchens 100 | accuracy: reduction of 1.62% at r=140 with 2.4x lower cost | Improved |
| apply Eventful conversion to ViViT | ViViT on Kinetics-400 · Kinetics-400 | top-1 accuracy: 79.06 -> 77.62; 3.360 -> 1.814 TFlops | Mixed |
| lower the policy budget r in Eventful ViTDet (1024) (ablation) | Eventful ViTDet · ImageNet VID val | mAP50: 82.94 (r=2048) -> 73.29 (r=256) | Hurt |
| add temporal redundancy awareness to ViTDet (GPU runtime) | ViTDet 1024 · GPU runtime | ms: 86.6 -> 69.9 (temporal); 86.6 -> 38.1 (spatiotemporal) | Improved |
Closest papers by meaning
- Recurrent Vision Transformers for Object Detection With Event Cameras (CVPR 2023 · similarity 0.90)
- IA-RED$^2$: Interpretability-Aware Redundancy Reduction for Vision Transformers (NeurIPS 2021 · similarity 0.89)
- Video Transformers: A Survey (TPAMI 2023 · similarity 0.89)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.87)
- Understanding Video Transformers via Universal Concept Discovery (CVPR 2024 · similarity 0.87)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.87)
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.
