Research map / Video / Action recognition
Leveraging Action Affinity and Continuity for Semi-Supervised Temporal Action Segmentation
Guodong Ding, Angela Yao
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 affinity and continuity losses plus ABS on unlabelled videos | MS-TCN base · 50Salads (5% labelled) | frame accuracy: 28.1 -> 52.3 | Improved |
| add naive pseudo-labelling on unlabelled videos | MS-TCN base · 50Salads (5% labelled) | frame accuracy: 28.1 -> 24.0 | Hurt |
| add naive pseudo-labelling on unlabelled videos | MS-TCN base · 50Salads (10% labelled) | frame accuracy: 32.0 -> 41.1 | Improved |
| use action affinity loss with entropy loss instead of pseudo-labelling (ablation) | semi-supervised MS-TCN · Breakfast | F1 and accuracy: around 2% | Improved |
| restrict anchor search to videos with the same activity label (ablation) | affinity loss · Breakfast (5% labelled) | frame accuracy: 38.1 -> 56.6 | Improved |
| increase sub-sampling stride from 20 to 60 (ablation) | action continuity loss · Breakfast (10% labelled) | frame accuracy: 49.5 -> 45.6 | Hurt |
| increase affinity loss weight alpha from 0.001 to 0.1 (ablation) | semi-supervised MS-TCN · GTEA (10% labelled) | frame accuracy: 58.0 -> 62.6 | Improved |
| raise continuity loss weight beta from 0.01 to 1 (ablation) | semi-supervised MS-TCN · GTEA (10% labelled) | frame accuracy: 62.6 -> 59.3 | Hurt |
| add ABS with v=0.05 over one-hot boundaries (ablation) | semi-supervised MS-TCN · Breakfast (10% labelled) | F1@10: 53.8 -> 56.9 | Improved |
| use fixed-duration linear smoothing instead of one-hot (ablation) | semi-supervised MS-TCN · Breakfast (10% labelled) | frame accuracy: 49.2 -> 49.1 | No clear effect |
| double vicinity from v=0.05 to 0.1 (ablation) | ABS · Breakfast (10% labelled) | frame accuracy: 49.5 -> 48.9 | Hurt |
| add ABS to fully-supervised MS-TCN | MS-TCN · Breakfast (fully supervised) | frame accuracy: 65.1 -> 68.9 | Improved |
Closest papers by meaning
- C2F-TCN: A Framework for Semi- and Fully-Supervised Temporal Action Segmentation (TPAMI 2023 · similarity 0.91)
- Iterative Contrast-Classify for Semi-supervised Temporal Action Segmentation (AAAI 2022 · similarity 0.91)
- Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment (CVPR 2024 · similarity 0.90)
- Unsupervised Action Segmentation via Fast Learning of Semantically Consistent Actoms (AAAI 2024 · similarity 0.89)
- OnlineTAS: An Online Baseline for Temporal Action Segmentation (NeurIPS 2024 · similarity 0.89)
- ActFusion: a Unified Diffusion Model for Action Segmentation and Anticipation (NeurIPS 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.
