Research map / Video / Action recognition
LogFA: Efficient Feature-Space Data Augmentation for Egocentric Temporal Action Segmentation
Zijia Lu, Ehsan Elhamifar
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 local text-guided augmentation (all four caption types) (ablation) | FACT · EgoPER Coffee (3 training videos) | F1@50: 32.2 -> 38.4 | Improved |
| add type3 texture/color caption changes vs type1 rephrasing (ablation) | FACT · EgoPER Coffee (3 training videos) | F1@50: 34.1 -> 36.8 | Improved |
| switch VLM backbone from SigLIP2 to EgoVLP (ablation) | LogFA · EgoPER Coffee (3 training videos) | F1@50: 38.4 -> 36.7 | Hurt |
| add dropout layers to base model (ablation) | FACT · EgoPER Coffee (3 training videos) | F1@50: 30.2 -> 32.2 | Improved |
| use GDAG instead of DAG for global augmentation (ablation) | FACT · EgoProceL (3 training videos), GDAG study | F1@50: 28.8 -> 32.2 | Improved |
| generate GDAG with Qwen3-30B instead of GPT-5o (ablation) | LogFA · EgoProceL (3 training videos), GDAG study | F1@50: 32.2 -> 31.7 | No clear effect |
| add LogFA (local + global) | FACT · EgoPER in-domain videos | F1@50: 64.1 -> 70.7 | Improved |
| add LogFA (local + global) | FACT · EgoPER out-of-domain videos | F1@50: 47.7 -> 57.0 | Improved |
| image transformation augmentation | FACT · EgoPER out-of-domain videos | F1@50: 47.7 -> 50.0 | Improved |
| add LogFA with MS-TCN++ backbone | MS-TCN++ · EgoPER all tasks (MS-TCN++) | F1@50: 54.5 -> 63.3 | Improved |
| add LogFA on EgoProceL | FACT · EgoProceL (3 training videos) | F1@50: 28.8 -> 35.0 | Improved |
| add LogFA with all training videos | FACT · EgoPER (all training videos) | F1@50: 63.1 -> 74.9 | Improved |
Closest papers by meaning
- Towards Generalizing Temporal Action Segmentation to Unseen Views (IJCV 2026 · similarity 0.89)
- Progress-Aware Online Action Segmentation for Egocentric Procedural Task Videos (CVPR 2024 · similarity 0.87)
- Action Scene Graphs for Long-Form Understanding of Egocentric Videos (CVPR 2024 · similarity 0.87)
- EgoVid-5M: A Large-Scale Video-Action Dataset for Egocentric Videos Generation (NeurIPS 2025 · similarity 0.87)
- Understanding Multi-Task Activities from Single-Task Videos (CVPR 2025 · similarity 0.87)
- Robust Egocentric Referring Video Object Segmentation via Dual-Modal Causal Intervention (NeurIPS 2025 · similarity 0.86)
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.
