Research map / Video / Video-language understanding
On the Hidden Treasure of Dialog in Video Question Answering
Deniz Engin, François Schnitzler, Ngoc Q. K. Duong, Yannis Avrithis
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 |
|---|---|---|---|
| replace human plot with generated episode dialog summary (ablation) | ROLL pipeline · KnowIT VQA test | accuracy (overall): 0.724 -> 0.775 | Improved |
| add scene dialog summary in place of dialog stream (ablation) | multi-stream QA · KnowIT VQA test | accuracy (overall): 0.775 -> 0.781 | Improved |
| add temporal attention (ablation) | ROLL pipeline · KnowIT VQA test | accuracy (overall): 0.715 -> 0.724 | Improved |
| increase plot parts from 5 to 10 (ablation) | ROLL pipeline · KnowIT VQA test | accuracy (overall): 0.703 -> 0.714 | Improved |
| change BERT argument order for episode inputs (ablation) | ROLL pipeline · KnowIT VQA test | accuracy (overall): 0.714 -> 0.715 | No clear effect |
| replace modality weighting with multi-stream attention (ablation) | ROLL pipeline · KnowIT VQA test | accuracy (overall): 0.704 -> 0.703 | No clear effect |
| add soft temporal attention to episode dialog summary single-stream (ablation) | single-stream QA · KnowIT VQA test (single-stream) | accuracy (overall): 0.723 -> 0.756 | Improved |
| add soft temporal attention to plot single-stream (ablation) | single-stream QA · KnowIT VQA test (single-stream) | accuracy (overall): 0.683 -> 0.702 | Improved |
| add soft temporal attention in single-stream for multi-stream V+S+E (ablation) | multi-stream attention · KnowIT VQA test | accuracy (overall): 0.711 -> 0.781 | Improved |
| use multi-stream attention instead of product fusion (ablation) | fusion V+S+E · KnowIT VQA test | accuracy (overall): 0.739 -> 0.781 | Improved |
| add self-attention to multi-stream attention (ablation) | multi-stream attention · KnowIT VQA test | accuracy (overall): 0.781 -> 0.770 | Hurt |
| use learned multi-stream attention instead of oracle question-type attention (ablation) | multi-stream attention · KnowIT VQA test | accuracy (overall): 0.745 -> 0.781 | Improved |
Closest papers by meaning
- Knowledge-Based Video Question Answering with Unsupervised Scene Descriptions (ECCV 2020 · similarity 0.90)
- LvBench: A Benchmark for Long-form Video Understanding with Versatile Multi-modal Question Answering (IJCV 2025 · similarity 0.89)
- KnowIT VQA: Answering Knowledge-Based Questions about Videos (AAAI 2020 · similarity 0.89)
- StoryVideoQA: Scaling Deep Video Understanding with a Large-Scale, Multi-Genre and Auto-Generated Dataset (IJCV 2026 · similarity 0.89)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.88)
- FriendsQA: A New Large-Scale Deep Video Understanding Dataset with Fine-grained Topic Categorization for Story Videos (AAAI 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.
