Research map / Vision-language and multimodal / Visual question answering
Zero-Shot Video Question Answering via Frozen Bidirectional Language Models
Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, Cordelia Schmid
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 pretrained language model instead of random init (ablation) | FrozenBiLM · MSVD-QA zero-shot | accuracy (%): 0.0 -> 31.9 | Improved |
| freeze language model weights (ablation) | FrozenBiLM · MSVD-QA zero-shot | accuracy (%): 31.9 -> 32.2 | Improved |
| freeze language model weights (ablation) | FrozenBiLM · How2QA zero-shot | accuracy (%): 45.7 -> 53.5 | Improved |
| freeze language model weights (ablation) | FrozenBiLM · MSRVTT-QA zero-shot | accuracy (%): 17.6 -> 16.8 | Hurt |
| add adapters to frozen language model (ablation) | FrozenBiLM · How2QA zero-shot | accuracy (%): 53.5 -> 58.4 | Improved |
| add adapters to frozen language model (ablation) | FrozenBiLM · TVQA zero-shot | accuracy (%): 53.4 -> 59.2 | Improved |
| add adapters to frozen language model (ablation) | FrozenBiLM · iVQA zero-shot | accuracy (%): 27.3 -> 26.8 | Hurt |
| add visual input (ablation) | FrozenBiLM · iVQA zero-shot | accuracy (%): 13.2 -> 26.8 | Improved |
| add speech input to model with vision (ablation) | FrozenBiLM · TVQA zero-shot | accuracy (%): 29.7 -> 59.2 | Improved |
| increase WebVid training data from 1K to 10M (ablation) | FrozenBiLM · MSVD-QA zero-shot | accuracy (%): 13.6 -> 33.8 | Improved |
| replace autoregressive GPT-J-6B with bidirectional DeBERTa-V2-XLarge (ablation) | FrozenBiLM · MSVD-QA zero-shot | accuracy (%): 26.7 -> 32.2 | Improved |
| increase bidirectional LM size from BERT-Base to DeBERTa-V2-XLarge (ablation) | FrozenBiLM · MSVD-QA zero-shot | accuracy (%): 11.7 -> 32.2 | Improved |
Closest papers by meaning
- From Images to Textual Prompts: Zero-Shot Visual Question Answering With Frozen Large Language Models (CVPR 2023 · similarity 0.92)
- Open-vocabulary Video Question Answering: A New Benchmark for Evaluating the Generalizability of Video Question Answering Models (ICCV 2023 · similarity 0.91)
- Learning to Answer Visual Questions From Web Videos (TPAMI 2022 · similarity 0.91)
- Just Ask: Learning To Answer Questions From Millions of Narrated Videos (ICCV 2021 · similarity 0.91)
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · similarity 0.89)
- Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning (CVPR 2025 · similarity 0.89)
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.
