Research map / Video / Action recognition
Temporal Aggregate Representations for Long-Range Video Understanding
Fadime Sener, Dipika Singhania, 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 |
|---|---|---|---|
| use max-pooling instead of frame sampling for snippets (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 32.1 -> 40.1 | Improved |
| use max-pooling instead of GRU (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 37.9 -> 40.1 | Improved |
| use max-pooling instead of mean-pooling (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 36.6 -> 40.1 | Improved |
| use non-local blocks instead of concatenation + linear layer (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 33.7 -> 40.1 | Improved |
| couple recent and spanning features in coupling blocks instead of only spanning or only recent (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 35.1 -> 40.1 | Improved |
| add Temporal Aggregation Blocks on top of a single coupling block (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 32.1 -> 40.1 | Improved |
| ensemble of recent starting points instead of single start (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 37.7 -> 40.1 | Improved |
| ensemble of three spanning scales instead of single scale (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 38.0 -> 40.1 | Improved |
| add fourth spanning scale (5,10,15,20) (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 40.1 -> 40.2 | No clear effect |
| use recent scale K_R=5 instead of 10 (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 38.6 -> 40.1 | Improved |
| add auxiliary complex activity prediction (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 1.1% | Improved |
| max over spanning outputs instead of concatenation + linear (ablation) | Temporal Aggregate model · Breakfast next action anticipation | accuracy: 37.4 -> 40.1 | Improved |
Closest papers by meaning
- Rethinking Learning Approaches for Long-Term Action Anticipation (ECCV 2022 · similarity 0.89)
- Learning Temporal Dynamics From Cycles in Narrated Video (ICCV 2021 · similarity 0.87)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.86)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.85)
- AntGPT: Can Large Language Models Help Long-term Action Anticipation from Videos? (ICLR 2024 · similarity 0.85)
- ActFusion: a Unified Diffusion Model for Action Segmentation and Anticipation (NeurIPS 2024 · similarity 0.85)
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.
