Research map / Video / Action recognition
Re2TAL: Rewiring Pretrained Video Backbones for Reversible Temporal Action Localization
Chen Zhao, Shuming Liu, Karttikeya Mangalam, Bernard Ghanem
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 |
|---|---|---|---|
| Train end to end instead of training on pre-extracted features (ablation) | Re²Vswin-tiny with VSGN · ActivityNet-v1.3 | average mAP (%): 34.47 -> 36.38 | Improved |
| Train end to end instead of training on pre-extracted features (ablation) | Re²Slowfast-101 with VSGN · ActivityNet-v1.3 | average mAP (%): 35.24 -> 36.82 | Improved |
| Add Kinetics-400-pretrained initialization rather than training from scratch (ablation) | Re²Vswin-Base · ActivityNet-v1.3 | average mAP (%): 28.68 -> 35.73 | Improved |
| Use Kinetics-600 rather than Kinetics-400 pretrained initialization (ablation) | Re²Vswin-Base · ActivityNet-v1.3 | average mAP (%): 35.73 -> 36.28 | Improved |
| Replace VSGN with ActionFormer (ablation) | Re²Vswin-tiny · ActivityNet-v1.3 | average mAP (%): 36.38 -> 36.80 | Improved |
| Replace VSGN with ActionFormer (ablation) | Re²Slowfast-101 · ActivityNet-v1.3 | average mAP (%): 36.82 -> 37.01 | Improved |
| Rewire the pretrained backbone and reuse its weights instead of using the original backbone (ablation) | Vswin-tiny feature extraction with VSGN · ActivityNet-v1.3 | average mAP (%): 34.36 -> 34.47 | Improved |
| Rewire the pretrained backbone and reuse its weights instead of using the original backbone (ablation) | Slowfast-101 feature extraction with VSGN · ActivityNet-v1.3 | average mAP (%): 35.04 -> 35.24 | Improved |
| Add spatial cropping, flipping, temporal jittering, rotation, and color jittering (ablation) | Re²Vswin-tiny · ActivityNet-v1.3 | average mAP: 35.25 -> 36.38 | Improved |
| Add random spatial cropping and flipping (ablation) | Re²Vswin-tiny · ActivityNet-v1.3 | average mAP: 35.25 -> 35.87 | Improved |
| Add temporal jittering to spatial cropping and flipping (ablation) | Re²Vswin-tiny · ActivityNet-v1.3 | average mAP: 35.87 -> 35.98 | Improved |
| Add random rotation to spatial cropping and flipping (ablation) | Re²Vswin-tiny · ActivityNet-v1.3 | average mAP: 35.87 -> 36.20 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ActionFormer — ActionFormer: Localizing Moments of Actions with Transformers (ECCV 2022) · better on 2 of 2 shared results
- TALLFormer — TALLFormer: Temporal Action Localization with a Long-Memory Transformer (ECCV 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Rethinking the Faster R-CNN Architecture for Temporal Action Localization (CVPR 2018 · similarity 0.89)
- Low-Fidelity Video Encoder Optimization for Temporal Action Localization (NeurIPS 2021 · similarity 0.89)
- Temporal Action Localization with Cross Layer Task Decoupling and Refinement (AAAI 2025 · similarity 0.88)
- Test-Time Zero-Shot Temporal Action Localization (CVPR 2024 · similarity 0.87)
- Masked Diffusion Vision-Language Models for Temporal Action Localization (NeurIPS 2026 · similarity 0.87)
- Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer (CVPR 2025 · 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.
