Research map / Architectures and efficiency / Attention and transformers
Normalized and Geometry-Aware Self-Attention Network for Image Captioning
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 |
|---|---|---|---|
| Add NSA and query-dependent GSA to the SAN encoder | SAN · MS-COCO Karpathy test | CIDEr: 128.6 -> 132.1 | Helped |
| Add NSA to the Transformer encoder | Transformer · VATEX | CIDEr: 53.4 -> 57.1 | Helped |
| Add NSA to the Transformer-Base encoder | Transformer-Base · WMT 2014 En-De newstest2014 | BLEU: 27.56 -> 27.92 | Helped |
| Replace MCAN self-attention with GSA | MCAN · VQA-v2 test-std | accuracy: 70.83 -> 71.28 | Helped |
| Replace SAN attention with NSA (ablation) | SAN, 4 layers · MS-COCO Karpathy test | CIDEr: 128.6 -> 130.8 | Helped |
| Use Layer Normalization instead of Instance Normalization inside attention (ablation) | NSA · MS-COCO Karpathy test | CIDEr: 128.6 -> 128.2 | Hurt |
| Use Batch Normalization inside attention (ablation) | NSA · MS-COCO Karpathy test | CIDEr: 128.6 -> 129.4 | Helped |
| Omit the affine parameters gamma and beta from Instance Normalization (ablation) | NSA · MS-COCO Karpathy test | CIDEr: 130.7 -> 130.8 | No clear effect |
| Normalize queries in addition to keys (ablation) | N-SAN · MS-COCO Karpathy test | CIDEr: 130.1 -> 130.7 | No clear effect |
| Normalize keys instead of queries (ablation) | N-SAN · MS-COCO Karpathy test | CIDEr: 130.8 -> 130.1 | Hurt |
| Add query-dependent relative geometry bias (ablation) | SAN · MS-COCO Karpathy test | CIDEr: 128.6 -> 131.4 | Helped |
| Use absolute geometry encodings instead of relative geometry bias (ablation) | SAN · MS-COCO Karpathy test | CIDEr: 128.6 -> 128.4 | Hurt |
Closest papers by meaning
- SCA-CNN: Spatial and Channel-Wise Attention in Convolutional Networks for Image Captioning (CVPR 2017 · similarity 0.88)
- Text-Guided Attention Model for Image Captioning (AAAI 2017 · similarity 0.88)
- Deliberate Attention Networks for Image Captioning (AAAI 2019 · similarity 0.88)
- Bottom-Up and Top-Down Attention for Image Captioning and Visual Question Answering (CVPR 2018 · similarity 0.88)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.86)
- MM-SeR: Multimodal Self-Refinement for Lightweight Image Captioning (CVPR 2026 · similarity 0.86)
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.