Research map / Vision-language and multimodal / Visual question answering
Improved Fusion of Visual and Language Representations by Dense Symmetric Co-Attention for Visual Question Answering
Read the paperCheck 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 |
|---|---|---|---|
| Add symmetric co-attention rather than use one attention direction (ablation) | DCN · VQA 2.0 val Open-Ended | Accuracy: 60.95 -> 62.94; 62.63 -> 62.94 | Helped |
| Increase memory size from K=1 to K=3 (ablation) | DCN · VQA 2.0 val Open-Ended | Accuracy: 62.53 -> 62.94 | Helped |
| Use memory size K=5 instead of K=3 (ablation) | DCN · VQA 2.0 val Open-Ended | Accuracy: 62.94 -> 62.83 | Hurt |
| Use 4 parallel attention maps rather than 2 or 8 (ablation) | DCN · VQA 2.0 val Open-Ended | Accuracy: 62.82 -> 62.94; 62.81 -> 62.94 | Helped |
| Use 3 stacked dense co-attention layers rather than 1, 2, or 4 (ablation) | DCN · VQA 2.0 val Open-Ended | Accuracy: 62.43 -> 62.94; 62.82 -> 62.94; 62.67 -> 62.94 | Helped |
| Add attention in the answer prediction layer instead of average feature pooling (ablation) | DCN · VQA 2.0 val Open-Ended | Accuracy: 61.63 -> 62.94 | Helped |
| Add question-guided attention across image extraction layers instead of using only the last convolutional layer (ablation) | DCN · VQA 2.0 val Open-Ended | Accuracy: 62.39 -> 62.94 | Helped |
| Use CoVe-pretrained Bi-LSTM instead of randomly initialized Bi-LSTM (ablation) | DCN (16) · VQA 2.0 Test-dev | Overall accuracy: 66.87 -> 67.06 | Helped |
| Use CoVe-pretrained Bi-LSTM instead of randomly initialized Bi-LSTM (ablation) | DCN (17) · VQA 2.0 Test-dev | Overall accuracy: 66.72 -> 66.21 | Hurt |
| Use CoVe-pretrained Bi-LSTM instead of randomly initialized Bi-LSTM (ablation) | DCN (18) · VQA 2.0 Test-dev | Overall accuracy: 66.60 -> 66.31 | Hurt |
| Use DCN instead of MF-SIG-T3 | Visual question answering · VQA 1.0 Test-standard | Overall accuracy: 65.88 -> 67.02 | Helped |
| Use DCN instead of MF-SIG-T3 | Visual question answering · VQA 2.0 Test-dev | Overall accuracy: 64.73 -> 66.87 | Helped |
Closest papers by meaning
- Hierarchical Question-Image Co-Attention for Visual Question Answering (NeurIPS 2016 · similarity 0.93)
- Multi-Level Attention Networks for Visual Question Answering (CVPR 2017 · similarity 0.92)
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.92)
- Deep Modular Co-Attention Networks for Visual Question Answering (CVPR 2019 · similarity 0.92)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.89)
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · 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.