Research map / Video / Action recognition
How Much Temporal Long-Term Context is Needed for Action Segmentation?
Emad Bahrami, Gianpiero Francesca, Juergen Gall
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 |
|---|---|---|---|
| combine windowed and long-term context attention in each block (ablation) | LTContext · 50Salads | F1@50: 77.4 -> 82.0 (windowed only); 74.1 -> 82.0 (LTC only) | Improved |
| increase window size W from 8 to 64 (ablation) | LTContext · 50Salads | F1@50: 74.5 -> 82.0 | Improved |
| decrease G (denser sparse attention) from 64 to 8 (ablation) | LTContext · 50Salads | F1@50: 82.0 -> 79.9 | Hurt |
| add overlap (2 windows) for keys and values in windowed attention (ablation) | LTContext · 50Salads | F1@50: 79.1 -> 82.0 | Improved |
| extend overlap to 3 windows (ablation) | LTContext · 50Salads | F1@50: 82.0 -> 80.8 | No clear effect |
| apply windowed attention before long-term context attention (ablation) | LTContext · 50Salads | F1@50: 80.4 -> 82.0 | Improved |
| add cross-attention using predictions in stages 2-4 (ablation) | LTContext · 50Salads | F1@50: 70.4 -> 82.0 | Improved |
| add dilated 1D convolution in the block (ablation) | LTContext · 50Salads | F1@50: 67.1 -> 82.0 (no conv); 67.7 -> 82.0 (no dilation) | Improved |
| use multiple attention heads instead of one (ablation) | LTContext · 50Salads | F1@50: 82.0 -> 80.5 (2 heads) | No clear effect |
| increase number of stages from 1 to 4 (ablation) | LTContext · 50Salads | F1@50: 49.9 -> 82.0 | Improved |
| increase number of stages from 4 to 5 (ablation) | LTContext · 50Salads | F1@50: 82.0 -> 80.0 | Hurt |
| increase layers per stage from 9 to 10 (ablation) | LTContext · mixed datasets | F1@50: 50Salads 82.0 -> 82.4; Breakfast 60.1 -> 59.7; Assembly101 22.6 -> 21.9 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- C2F-TCN — C2F-TCN: A Framework for Semi- and Fully-Supervised Temporal Action Segmentation (TPAMI 2023) · better on 1 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
- OnlineTAS: An Online Baseline for Temporal Action Segmentation (NeurIPS 2024 · similarity 0.90)
- Temporal Convolutional Networks for Action Segmentation and Detection (CVPR 2017 · similarity 0.89)
- Temporal Segment Networks for Action Recognition in Videos (TPAMI 2018 · similarity 0.89)
- SCT: Set Constrained Temporal Transformer for Set Supervised Action Segmentation (CVPR 2020 · similarity 0.89)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.88)
- FACT: Frame-Action Cross-Attention Temporal Modeling for Efficient Action Segmentation (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.
