Research map / Video / Video-language understanding
Self-Chained Image-Language Model for Video Localization and Question Answering
Read the paperCheck 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 uniform 4-frame sampling with zero-shot language-aware keyframe selection (ablation) | BLIP-2 Answerer · NExT-QA zero-shot | accuracy (Avg.): 62.4 -> 63.6 | Helped |
| Refine the Localizer using Answerer-generated pseudo-labels (ablation) | SeViLA · NExT-QA zero-shot | accuracy (Avg.): 63.6 -> 64.6 | Helped |
| Use 4 rather than 32 uniformly sampled frames (ablation) | BLIP-2 Answerer · NExT-QA zero-shot | accuracy (Avg.): 57.7 -> 62.4 | Helped |
| Add QVHighlights pre-training to the Localizer (ablation) | SeViLA Localizer · NExT-QA zero-shot | accuracy (Avg.): 62.9 -> 63.6 | Helped |
| Add weakly supervised QVH ASR pre-training to the Localizer (ablation) | SeViLA Localizer · NExT-QA zero-shot | accuracy (Avg.): 62.9 -> 63.2 | Helped |
| Use the Localizer for keyframe selection rather than uniform sampling (ablation) | BLIP-2 Answerer · NExT-QA fine-tuning | accuracy (Avg.): 72.6 -> 73.4 | Helped |
| Refine the Localizer with pseudo-labels (ablation) | SeViLA · NExT-QA fine-tuning | accuracy (Avg.): 73.4 -> 73.8 | Helped |
| Increase self-refinement from one to two iterations (ablation) | SeViLA · NExT-QA fine-tuning | accuracy (Avg.): 73.8 -> 74.2 | Helped |
| Increase self-refinement from two to three iterations (ablation) | SeViLA · NExT-QA fine-tuning | accuracy (Avg.): 74.2 -> 73.7 | Hurt |
| Use a multi-frame 4-frame Localizer instead of a single-frame Localizer (ablation) | SeViLA Localizer · NExT-QA zero-shot | accuracy (Avg.): 64.6 -> 63.6 | Hurt |
| Use a multi-frame 4-frame Localizer instead of a single-frame Localizer (ablation) | SeViLA Localizer · NExT-QA fine-tuning | accuracy (Avg.): 73.4 -> 71.3 | Hurt |
| Use Localizer-selected rather than uniformly sampled frames | BLIP-2 Answerer · QVHighlights test | R1@0.5 | Mixed |
Closest papers by meaning
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · similarity 0.88)
- Video-Panda: Parameter-efficient Alignment for Encoder-free Video-Language Models (CVPR 2025 · similarity 0.88)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.88)
- Omnipotent Distillation with LLMs for Weakly-Supervised Natural Language Video Localization: When Divergence Meets Consistency (AAAI 2024 · similarity 0.88)
- Localizing Events in Videos with Multimodal Queries (CVPR 2025 · similarity 0.88)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · 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.