Research map / Video / Video segmentation and tracking
(2.5+1)D Spatio-Temporal Scene Graphs for Video Question Answering
Anoop Cherian, Chiori Hori, Tim K. Marks, Jonathan Le Roux
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 the (2.5+1)D-Transformer instead of graph-alignment VQA | video QA · NExT-QA validation | accuracy (%): 49.74 -> 53.4 | Improved |
| use the (2.5+1)D-Transformer instead of ST Scene Graphs | video QA · AVSD-QA test | mean rank: 5.91 -> 5.84 | Improved |
| remove the dynamic graph (ablation) | full (2.5+1)D-Transformer · NExT-QA validation | accuracy (%): 52.49 -> 53.4 | Hurt |
| remove the static graph (ablation) | full (2.5+1)D-Transformer · NExT-QA validation | accuracy (%): 53.0 -> 53.4 | No clear effect |
| remove I3D motion features (ablation) | full (2.5+1)D-Transformer · NExT-QA validation | accuracy (%): 52.65 -> 53.4 | Hurt |
| remove augmented-answer training (ablation) | full (2.5+1)D-Transformer · NExT-QA validation | accuracy (%): 49.98 -> 53.4 | Hurt |
| remove question conditioning (ablation) | full (2.5+1)D-Transformer · NExT-QA validation | accuracy (%): 50.39 -> 53.4 | Hurt |
| remove the hierarchical kernel (ablation) | full (2.5+1)D-Transformer · NExT-QA validation | accuracy (%): 52.9 -> 53.4 | Hurt |
| add pseudo-depth to the no-depth V(2+1)D Transformer (ablation) | NExT-QA video QA · NExT-QA validation | accuracy (%): 52.4 -> 53.4 | Improved |
| use all graph nodes instead of pruning (ablation) | (2.5+1)D-Transformer · NExT-QA validation | accuracy (%): 53.4 -> 53.5 | Improved |
| increase hierarchy levels from 1 to 4 using the listed bandwidths (ablation) | (2.5+1)D-Transformer · NExT-QA validation | accuracy: 52.13 -> 53.20 | Improved |
| increase hierarchy levels from 4 to 5 (ablation) | (2.5+1)D-Transformer · NExT-QA validation | accuracy: 53.20 -> 53.0 | Hurt |
Closest papers by meaning
- Location-Aware Graph Convolutional Networks for Video Question Answering (AAAI 2020 · similarity 0.92)
- AGQA: A Benchmark for Compositional Spatio-Temporal Reasoning (CVPR 2021 · similarity 0.90)
- Tool-Augmented Spatiotemporal Reasoning for Streamlining Video Question Answering Task (NeurIPS 2025 · similarity 0.90)
- STEP: Enhancing Video-LLMs' Compositional Reasoning by Spatio-Temporal Graph-guided Self-Training (CVPR 2025 · similarity 0.90)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.89)
- Compositional 4D Dynamic Scenes Understanding with Physics Priors for Video Question Answering (ICLR 2025 · similarity 0.88)
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.
