Research map / Video / Action recognition
Future Transformer for Long-Term Action Anticipation
Check 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 |
|---|---|---|---|
| Replace autoregressive label decoding with parallel action queries (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 20.51 -> 24.17 | Helped |
| Remove the causal mask so action queries can attend bidirectionally (ablation) | FUTR-M on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 24.17 -> 25.87 | Helped |
| Replace local attention with global attention in both encoder and decoder (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 21.60 -> 25.87 | Helped |
| Add global attention in the encoder while retaining local decoder attention (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 21.60 -> 23.28 | Helped |
| Add global attention in the decoder while retaining local encoder attention (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 21.60 -> 22.28 | Helped |
| Use sequential ground-truth assignment and duration regression rather than start-end regression (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 21.43 -> 25.87 | Helped |
| Use sequential assignment rather than Hungarian assignment with start-end regression (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 21.45 -> 21.43 | Helped |
| Add action segmentation loss for observed frames (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 22.50 -> 25.87 | Helped |
| Allow decoder cross-attention to use all observed frames rather than only the recent 0.25 (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 20.10 -> 25.87 | Helped |
| Use learnable positional embeddings at attention layers rather than no positional embeddings (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 18.25 -> 25.87 | Helped |
| Use 2 encoder layers and 1 decoder layer rather than 3 encoder layers and 1 decoder layer (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 23.75 -> 25.87 | Helped |
| Set hidden dimension to 128 rather than 64, 256, or 512 (ablation) | FUTR on Breakfast · Breakfast (α=0.3, β=0.5) | MoC accuracy: 23.53, 24.29, 22.56 -> 25.87 | Helped |
Closest papers by meaning
- AntGPT: Can Large Language Models Help Long-term Action Anticipation from Videos? (ICLR 2024 · similarity 0.90)
- Towards Understanding Future: Consistency Guided Probabilistic Modeling for Action Anticipation (AAAI 2024 · similarity 0.89)
- Can't Make an Omelette Without Breaking Some Eggs: Plausible Action Anticipation Using Large Video-Language Models (CVPR 2024 · similarity 0.88)
- Uncertainty-aware Action Decoupling Transformer for Action Anticipation (CVPR 2024 · similarity 0.88)
- PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning (ICML 2025 · similarity 0.86)
- MANTA: Diffusion Mamba for Efficient and Effective Stochastic Long-Term Dense Action Anticipation (CVPR 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.