Research map / Vision-language and multimodal / Visual question answering
TAB-VCR: Tags and Attributes based VCR Baselines
Jingxiang Lin, Unnat Jain, Alexander Schwing
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 Det-BN (train-time batch norm statistics) (ablation) | R2C · VCR val | Q->AR accuracy: 43.1 -> 43.61 | Improved |
| freeze image CNN weights (ablation) | R2C + Det-BN · VCR val | Q->AR accuracy: 43.61 -> 44.41 | Improved |
| use ResNet101 backbone (ablation) | R2C++ · VCR val | Q->AR accuracy: 44.41 -> 46.42 | Improved |
| add attribute features (ablation) | R2C++ + ResNet101 · VCR val | Q->AR accuracy: 46.42 -> 48.64 | Improved |
| use ResNet101 instead of ResNet50 (ablation) | Base · VCR val | Q->AR accuracy: 46.19 -> 47.51 | Improved |
| add attributes (ablation) | Base + ResNet101 · VCR val | Q->AR accuracy: 47.51 -> 50.08 | Improved |
| add new tags for ungrounded nouns (ablation) | Base + ResNet101 + Attributes · VCR val | Q->AR accuracy: 50.08 -> 50.62 | Improved |
| unshare encoder parameters between query and response (ablation) | TAB-VCR · VCR val | Q->AR accuracy: 50.62 -> 50.35 | No clear effect |
| freeze 4th conv block instead of fine-tuning (ablation) | Base · VCR val | Q->AR accuracy: 44.60 -> 46.45 | Improved |
| lower learning rate for fine-tuning conv block (0.5x, 0.25x, 0.125x) (ablation) | Base · VCR val | Q->AR accuracy: 44.60 -> 44.08 | Hurt |
| remove downsample net (ablation) | Base (frozen conv block) · VCR val | Q->AR accuracy: 46.45 -> 45.57 | Hurt |
| more image-text grounding (tags) in query+response | TAB-VCR · VCR val | Q->A tags: 2.566 -> 2.719 | Improved |
Closest papers by meaning
- A Case Study of the Shortcut Effects in Visual Commonsense Reasoning (AAAI 2021 · similarity 0.90)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.88)
- VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning (CVPR 2024 · similarity 0.87)
- VCR: A Task for Pixel-Level Complex Reasoning in Vision Language Models via Restoring Occluded Text (ICLR 2025 · similarity 0.87)
- Not Another Text Benchmark: Putting the “Visual" Back in Visual Question Answering for Large Video Models (NeurIPS 2026 · similarity 0.87)
- Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning (CVPR 2025 · similarity 0.87)
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.
