Research map / Vision-language and multimodal / Visual question answering
In Defense of Grid Features for Visual Question Answering
Check 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 region features with C5 grid features from the same detector layer | VQA co-attention model · VQA 2.0 valminusminival | accuracy: 64.29 -> 63.64 | No clear effect |
| Train grid features with 1x1 RoIPool and two FC layers (ablation) | Grid C5 VQA features · VQA 2.0 valminusminival | accuracy: 63.64 -> 64.37 | Helped |
| Use 1x1 RoIPool and two FC layers for region features (ablation) | BUTD regions · VQA 2.0 valminusminival | accuracy: 64.29 -> 63.94 | Hurt |
| Increase region feature count from 100 to 608 (ablation) | BUTD regions · VQA 2.0 test | accuracy: 66.13 -> 66.22 | No clear effect |
| Replace regions with grid features and skip region computation and selection | VQA co-attention model · VQA 2.0 test | total inference time (ms/image): 889 -> 18 | Helped |
| Increase input resolution (ablation) | VG-pretrained grid features · VQA 2.0 valminusminival | accuracy: 63.24 -> 63.81 -> 64.37 -> 64.61 | Helped |
| Increase input resolution from 600x1000 to 800x1333 (ablation) | ImageNet-pretrained grid features · VQA 2.0 valminusminival | accuracy: 61.52 -> 61.52 | No clear effect |
| Use YFCC image-tag pre-training instead of ImageNet classification pre-training (ablation) | Grid features · VQA 2.0 valminusminival | accuracy: 61.52 -> 62.72 | Helped |
| Use VG object-and-attribute detection pre-training instead of COCO object-box pre-training (ablation) | Grid features · VQA 2.0 valminusminival | accuracy: 62.46 -> 64.37 | Helped |
| Replace regions with grids | MCAN with ResNeXt-101-32x8d features · VQA 2.0 test | accuracy: 72.01 -> 72.59 | Helped |
| Replace regions with grids | Pythia with ResNeXt-101-32x8d features · VizWiz test | accuracy: 54.28 -> 54.17 | No clear effect |
| Fine-tune grid features end-to-end for the answer task (ablation) | ImageNet-pretrained grid model · VQA 2.0 test | accuracy: 63.21 -> 64.98 | Helped |
Closest papers by meaning
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.92)
- Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering (CVPR 2018 · similarity 0.91)
- REVIVE: Regional Visual Representation Matters in Knowledge-Based Visual Question Answering (NeurIPS 2022 · similarity 0.91)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.90)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.89)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 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.