Research map / Architectures and efficiency / Attention and transformers
Efficient Attention Mechanism for Visual Dialog that can Handle All the Interactions between Multiple Inputs
Van-Quang Nguyen, Masanori Suganuma, Takayuki Okatani
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 self-attention within each utility (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 61.76 -> 62.72 | Improved |
| concatenate attended features instead of adding them (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 60.28 -> 62.72 | Improved |
| do not share attention weights across blocks (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 62.50 -> 62.72 | Improved |
| use 2 attention blocks instead of 1 (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 62.06 -> 62.72 | Improved |
| use 3 blocks instead of 2 (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 62.72 -> 62.48 | No clear effect |
| add image context vector c_V to question context (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 61.50 -> 62.72 | Improved |
| add dialog-history context c_R (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 62.72 -> 62.37 | No clear effect |
| train both decoders jointly vs discriminative only (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 61.80 -> 62.72 | Improved |
| train both decoders jointly vs generative only (ablation) | LTMI · VisDial v1.0 val (ablation) | G-NDCG: 62.35 -> 63.58 | Improved |
| use 100 objects instead of 36 (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 62.40 -> 62.72 | Improved |
| add positional and spatial embeddings (ablation) | LTMI · VisDial v1.0 val (ablation) | D-NDCG: 61.84 -> 62.72 | Improved |
| replace naive Transformer extension with LTMI attention | LTMI · VisDial v1.0 test-standard | NDCG: 55.10 -> 60.92 | Improved |
Closest papers by meaning
- Visual Dialog (CVPR 2017 · similarity 0.90)
- Structure-Aware Multimodal Sequential Learning for Visual Dialog (AAAI 2024 · similarity 0.87)
- MMDU: A Multi-Turn Multi-Image Dialog Understanding Benchmark and Instruction-Tuning Dataset for LVLMs (NeurIPS 2024 · similarity 0.86)
- V^2Dial: Unification of Video and Visual Dialog via Multimodal Experts (CVPR 2025 · similarity 0.86)
- VISion On Request: Enhanced VLLM efficiency with sparse, dynamically selected, vision-language interactions (CVPR 2026 · similarity 0.85)
- Memorization Capacity of Multi-Head Attention in Transformers (ICLR 2024 · similarity 0.85)
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.
