Research map / Vision-language and multimodal / Visual question answering
Spatially Aware Multimodal Transformers for TextVQA
Yash Kant, Dhruv Batra, Peter Anderson, Alexander Schwing, Devi Parikh, Jiasen Lu, Harsh Agrawal
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 Rosetta-en OCR with Google OCR | M4C · TextVQA val | VQA accuracy: 39.4 -> 41.8 | Improved |
| replace ResNet-101 with ResNeXt-151 detector backbone | M4C · TextVQA val | VQA accuracy: 41.8 -> 42.0 | Improved |
| add two more transformer layers (6N) | M4C · TextVQA val | VQA accuracy: 42.0 -> 42.7 | Improved |
| jointly train on ST-VQA | M4C improved · TextVQA val | VQA accuracy: 42.7 -> 43.3 | Improved |
| use beam search (size 5) | M4C improved · TextVQA val | VQA accuracy: 43.3 -> 43.8 | Improved |
| replace 4 of 6 self-attention layers with spatially aware attention (ablation) | M4C improved · TextVQA val | VQA accuracy: 42.70 -> 43.80 | Improved |
| replace only 1 of 6 layers with spatially aware attention (ablation) | M4C improved · TextVQA val | VQA accuracy: 42.70 -> 42.61 | No clear effect |
| replace 5 of 6 layers (leaving one normal layer) instead of 4 (ablation) | SA-M4C · TextVQA val | VQA accuracy: 43.80 -> 43.07 | Hurt |
| set context size c=2 instead of 1 (ablation) | SA-M4C · TextVQA val | VQA accuracy: 43.80 -> 43.90 | Improved |
| use spatial graph masking instead of random masking (ablation) | SA-M4C · TextVQA val | VQA accuracy: 42.09 -> 43.80 | Improved |
| use spatial graph instead of top-k attention (ablation) | SA-M4C · TextVQA val | VQA accuracy: 43.26 -> 43.90 | Improved |
| use spatially aware attention instead of ReGAT graph attention (ablation) | SA-M4C · TextVQA val | VQA accuracy: 43.20 -> 43.90 | Improved |
Closest papers by meaning
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.89)
- Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas (ICML 2025 · similarity 0.89)
- SIFThinker: Spatially-Aware Image Focus for Visual Reasoning (AAAI 2026 · similarity 0.88)
- Enhancing Spatial Reasoning Through Visual and Textual Thinking (AAAI 2026 · similarity 0.88)
- SpatiaLQA: A Benchmark for Evaluating Spatial Logical Reasoning in Vision-Language Models (CVPR 2026 · similarity 0.88)
- Graph-of-Mark: Promote Spatial Reasoning in Multimodal Language Models with Graph-Based Visual Prompting (AAAI 2026 · 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.
