Research map / Video / Action recognition
EAST: Early Action Prediction Sampling Strategy with Token Masking
Iva Sović, Ivan Martinović, Marin Oršić
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 |
|---|---|---|---|
| train with randomly sampled observation ratio (EAST sampling) instead of standard classification finetuning (ablation) | VideoMAE ViT-B/16 · SSv2 (rho=0.1) | top-1 accuracy (%): 9.9 -> 23.9 | Improved |
| sampling strategy at high observation ratio (ablation) | VideoMAE ViT-B/16 · SSv2 (rho=0.9) | top-1 accuracy (%): 63.4 -> 60.7 | Hurt |
| add oracle-feature loss to encoder-only model (ablation) | EAST_E · SSv2 (average) | top-1 accuracy (%): 44.8 -> 46.3 | Improved |
| add oracle-feature loss to encoder-decoder model (ablation) | EAST · SSv2 (average) | top-1 accuracy (%): 46.2 -> 46.9 | Improved |
| add transformer decoder to encoder with both losses (ablation) | EAST_E · SSv2 (average) | top-1 accuracy (%): 46.3 -> 46.9 | Improved |
| use direct decoding instead of autoregressive decoding (ablation) | EAST · SSv2 (rho=0.5) | top-1 accuracy (%): 48.2 -> 49.0 | Improved |
| add L2 feature alignment loss (ablation) | EAST · SSv2 (rho=0.5) | top-1 accuracy (%): 49.0 -> 48.2 | Hurt |
| use shared classifier instead of separate heads (ablation) | EAST · SSv2 (rho=0.5) | top-1 accuracy (%): 48.3 -> 49.0 | Improved |
| use 4-layer decoder instead of 1 or 12 layers (ablation) | EAST · UCF101 (average) | top-1 accuracy (%): 94.1 -> 94.9 | Improved |
| use 4-layer decoder instead of 1 or 12 layers (ablation) | EAST · SSv2 (average) | top-1 accuracy (%): 46.8 -> 46.9 | No clear effect |
| train one model for all observation ratios instead of one per ratio (ablation) | EAST · all datasets | accuracy | No clear effect |
| difference token masking at k=0.5 for efficiency (ablation) | EAST · UCF101 training epoch | seconds per epoch: 180 -> 80 | Improved |
Closest papers by meaning
- Encouraging LSTMs to Anticipate Actions Very Early (ICCV 2017 · similarity 0.87)
- Can't Make an Omelette Without Breaking Some Eggs: Plausible Action Anticipation Using Large Video-Language Models (CVPR 2024 · similarity 0.84)
- Context-Enhanced Memory-Refined Transformer for Online Action Detection (CVPR 2025 · similarity 0.84)
- ActFusion: a Unified Diffusion Model for Action Segmentation and Anticipation (NeurIPS 2024 · similarity 0.84)
- VEDIT: Latent Prediction Architecture For Procedural Video Representation Learning (ICLR 2025 · similarity 0.84)
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · 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.
