Research map / Video / Action recognition
When Will You Do What? - Anticipating Temporal Occurrences of Activities
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| predict future segments recursively with the RNN rather than using the grammar baseline | RNN model · Breakfast and 50Salads, 20% observed; ground-truth observations | MoC accuracy: Breakfast: 0.4892 -> 0.6035 (10% prediction), 0.3146 -> 0.4042 (50%); 50Salads: 0.2869 -> 0.4230 (10%), 0.1037 -> 0.1682 | Helped |
| use the CNN one-pass predictor rather than the grammar baseline | CNN model · Breakfast and 50Salads, 20% observed; ground-truth observations | MoC accuracy: Breakfast: 0.4892 -> 0.5797 (10% prediction), 0.3146 -> 0.3926 (50%); 50Salads: 0.2869 -> 0.3608 (10%), 0.1037 -> 0.1548 | Helped |
| use RNN rather than CNN anticipation | RNN and CNN models · 20% observed, ground-truth observations | MoC accuracy: Breakfast: 0.6035 vs 0.5797 (10% prediction), 0.4042 vs 0.3926 (50%); 50Salads: 0.4230 vs 0.3608 (10%), 0.1682 vs 0.1548 | Mixed |
| use RNN rather than CNN anticipation | RNN and CNN models · Breakfast and 50Salads, 20% observed; noisy observations | MoC accuracy: Breakfast: 0.1811 vs 0.1790 (10% prediction), 0.1581 vs 0.1454 (50%); 50Salads: 0.3006 vs 0.2124 (10%), 0.1349 vs 0.0987 | Mixed |
| replace squared loss with cross-entropy in the CNN (ablation) | CNN model · Breakfast, ground-truth observations | MoC accuracy: 0.5797 -> 0.5177; 0.4912 -> 0.4280; 0.4403 -> 0.3940; 0.3926 -> 0.3625 | Hurt |
| predict directly from frame features instead of first decoding observed action labels (ablation) | CNN model · Breakfast, noisy observations | MoC accuracy: 0.1278 vs 0.1790; 0.1162 vs 0.1635; 0.1121 vs 0.1537; 0.1027 vs 0.1454; 0.1772 vs 0.2244; 0.1687 vs 0.2012; 0.1548 vs 0. | Hurt |
| use temporal context from the observed video rather than predicting from a single future-frame representation | RNN and CNN models · Breakfast and 50Salads, one second before a future action | accuracy: Breakfast: 8.1 -> 30.1 (RNN), 8.1 -> 27 (CNN); 50Salads: 6.2 -> 30.1 (RNN), 6.2 -> 29.8 (CNN) | Helped |
| use squared error with row-wise L2 normalization rather than cross-entropy (ablation) | CNN model · Breakfast, ground-truth observations | MoC accuracy: 0.5177 -> 0.5797; 0.4280 -> 0.4912; 0.3940 -> 0.4403; 0.3625 -> 0.3926 | Helped |
Closest papers by meaning
- Anticipating Visual Representations From Unlabeled Video (CVPR 2016 · similarity 0.90)
- Future Transformer for Long-Term Action Anticipation (CVPR 2022 · similarity 0.89)
- Time-Conditioned Action Anticipation in One Shot (CVPR 2019 · similarity 0.89)
- Towards Understanding Future: Consistency Guided Probabilistic Modeling for Action Anticipation (AAAI 2024 · similarity 0.85)
- AntGPT: Can Large Language Models Help Long-term Action Anticipation from Videos? (ICLR 2024 · similarity 0.85)
- FutureHuman3D: Forecasting Complex Long-Term 3D Human Behavior from Video Observations (CVPR 2024 · similarity 0.84)
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.