Research map / Video / Video-language understanding
We Have So Much In Common: Modeling Semantic Relational Set Abstractions in Videos
Alex Andonian, Camilo Fosco, Mathew Monfort, Allen Lee, Rogerio Feris, Carl Vondrick, Aude Oliva
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 set abstraction module (SAM) to the base model | 3DResNet50 · M-MiT recognition N=2 | top-1 accuracy: 17.1 -> 34.0 | Improved |
| add SAM | 3DResNet50 · M-MiT recognition N=4 | top-1 accuracy: 26.0 -> 47.2 | Improved |
| add SAM | 3DResNet50 · Kinetics recognition N=2 | top-1 accuracy: 29.9 -> 60.5 | Improved |
| add SAM | 3DResNet50 · Kinetics recognition N=4 | top-1 accuracy: 17.9 -> 69.9 | Improved |
| use SAM power-set module instead of Relation Network pairs (ablation) | 3DResNet50 · Kinetics recognition N=2 | top-1 accuracy: 53.9 -> 60.5 | Improved |
| use SAM power-set module instead of Relation Network pairs (ablation) | 3DResNet50 · M-MiT recognition N=4 | top-1 accuracy: 44.9 -> 47.2 | Improved |
| train with multi-label BCE over ancestor labels instead of standard classification | 3DResNet50 · M-MiT recognition N=2 | top-1 accuracy: 17.1 -> 3.9 | Hurt |
| add SAM | 3DResNet50 · M-MiT set completion | rank correlation (avg): 0.430 -> 0.542 | Improved |
| add SAM | 3DResNet50 · Kinetics set completion | rank correlation (avg): 0.413 -> 0.604 | Improved |
| add SAM | 3DResNet50 · M-MiT odd one out N=3 | top-1 accuracy: 49.95 -> 52.63 | Improved |
| add SAM | 3DResNet50 · Kinetics odd one out N=3 | top-1 accuracy: 65.15 -> 85.90 | Improved |
| add SAM | 3DResNet50 · M-MiT odd one out N=4 | top-1 accuracy: 43.73 -> 47.21 | Improved |
Closest papers by meaning
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.86)
- EventLens: Event-Structure Reinforcement Learning for Video Understanding (NeurIPS 2026 · similarity 0.86)
- ConViS-Bench: Estimating Video Similarity Through Semantic Concepts (NeurIPS 2025 · similarity 0.86)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 2025 · similarity 0.86)
- Set Prediction Guided by Semantic Concepts for Diverse Video Captioning (AAAI 2024 · similarity 0.86)
- OpenCoF: Learning to Reason Through Video Generation (NeurIPS 2026 · 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.
