Research map / Video / Action recognition
TALLFormer: Temporal Action Localization with a Long-Memory Transformer
Feng Cheng, Gedas Bertasius
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 consistency module (ablation) | TALLFormer · THUMOS14 (Swin-T 112) | average mAP: 51.1 -> 52.6 | Improved |
| increase temporal support from 8s to 40s (ablation) | TALLFormer · THUMOS14 (Swin-T 112) | average mAP: 41.5 -> 53.3 | Improved |
| reduce clip sampling rate from 0.6 to 0.15 (ablation) | TALLFormer · THUMOS14 (Swin-B 224) | average mAP: 60.0 -> 59.0 | No clear effect |
| use long memory with sampling instead of full end-to-end (r=1.0) (ablation) | TALLFormer · THUMOS14 (Swin-B 224) | GPU memory: 45 GB at r=0.6 -> 12 GB at r=0.15 | Improved |
| unfreeze backbone (end-to-end) instead of frozen features | TALLFormer vs LT-Frozen · THUMOS14 (12GB cap) | average mAP: 52.7 -> 59.0 | Improved |
| use long temporal support with memory instead of short temporal support ST-E2E | TALLFormer vs ST-E2E · THUMOS14 (32GB cap) | average mAP: 55.1 -> 59.2 | Improved |
| use dense 15 fps instead of sparse 8 fps for 32s ST-E2E | ST-E2E · THUMOS14 (12GB cap) | average mAP: 42.9 -> 31.9 | Hurt |
| use larger Swin backbone (Swin-B instead of Swin-T) (ablation) | TALLFormer · THUMOS14 | average mAP: 57.0 -> 59.2 | Improved |
| raise spatial resolution from 112x112 to 168x168 | DaoTAD Swin-T · THUMOS14 | mAP: 52.8 -> 55.1 | Improved |
| freeze 4 backbone stages instead of 2 | DaoTAD Swin-T · THUMOS14 | mAP: 52.8 -> 44.3 | Hurt |
| use Swin-B backbone instead of I3D | TALLFormer-12 · ActivityNet-1.3 | average mAP: 27.2 -> 33.7 | Improved |
| use single RGB model instead of two-stream with external classifier | TALLFormer-32 vs AFSD · ActivityNet-1.3 | average mAP: 34.4 -> 35.6 | Improved |
Later papers that compared with this method
3 later accepted papers report TALLFormer as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- SV-TAD: Native Sparse Convolutions for Efficient Temporal Action Detection (ECCV 2026 · better on 1 of 1 shared result)
- Boundary Denoising for Video Activity Localization (ICLR 2024 · better on 1 of 1 shared result)
- Re2TAL: Rewiring Pretrained Video Backbones for Reversible Temporal Action Localization (CVPR 2023 · better on 2 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GTAD — G-TAD: Sub-Graph Localization for Temporal Action Detection (CVPR 2020) · better on 1 of 1 shared result
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
- ActionFormer: Localizing Moments of Actions with Transformers (ECCV 2022 · similarity 0.92)
- Temporal Action Localization in the Deep Learning Era: A Survey (TPAMI 2023 · similarity 0.90)
- Online Temporal Action Localization with Memory-Augmented Transformer (ECCV 2024 · similarity 0.90)
- Temporal Action Localization in Untrimmed Videos via Multi-Stage CNNs (CVPR 2016 · similarity 0.90)
- Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer (CVPR 2025 · similarity 0.88)
- End-to-End Spatio-Temporal Action Localisation with Video Transformers (CVPR 2024 · similarity 0.88)
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.
