Research map / Vision-language and multimodal / Visual question answering
Learning to Collocate Visual-Linguistic Neural Modules for Image Captioning
Xu Yang, Hanwang Zhang, Chongyang Gao, Jianfei Cai
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 |
|---|---|---|---|
| fuse all four modules with equal weights instead of a single RELATION module (ablation) | Module/R · MS-COCO Karpathy test | CIDEr-D: 123.8 -> 125.6 | Improved |
| use soft learnable fusion weights instead of equal weights (ablation) | Col/1 · MS-COCO Karpathy test | CIDEr-D: 125.6 -> 127.3 | Improved |
| use soft fusion instead of hard Gumbel-Softmax selection (ablation) | Col/H · MS-COCO Karpathy test | CIDEr-D: 126.1 -> 127.3 | Improved |
| replace LSTM controller with multi-head attention controller (ablation) | LSTM · MS-COCO Karpathy test | CIDEr-D: 125.7 -> 126.7 | Improved |
| feed previous module layout Z into the MH-ATT controller (ablation) | MH-ATT · MS-COCO Karpathy test | CIDEr-D: 126.7 -> 127.3 | Improved |
| add part-of-speech syntax loss (ablation) | MH-ATT+Z · MS-COCO Karpathy test | CIDEr-D: 127.3 -> 128.4 | Improved |
| add part-of-speech syntax loss (ablation) | MH-ATT+Z · MS-COCO Karpathy test | module layout accuracy: 68.6 -> 77.6 | Improved |
| add Reason module (ablation) | MH-ATT+Z · MS-COCO Karpathy test | CIDEr-D: 127.3 -> 128.6 | Improved |
| add part-of-speech syntax loss (ablation) | MH-ATT+Z · MS-COCO Karpathy test | CHAIRi: 7.8 -> 7.4 | Improved |
| add Reason module on top of syntax loss (full CVLNM) (ablation) | MH-ATT+Z+Ls · MS-COCO Karpathy test | CIDEr-D: 128.4 -> 129.5 | Improved |
| change syntax-loss weight lambda to 5 (ablation) | CVLNM · MS-COCO Karpathy test | CIDEr-D: 129.5 -> 129.2 | No clear effect |
| pre-train CNN with 1600/400 objects/attributes instead of 305/103 (ablation) | CVLNM · MS-COCO Karpathy test | CIDEr-D: 129.5 -> 130.1 | Improved |
Closest papers by meaning
- Learning to Collocate Neural Modules for Image Captioning (ICCV 2019 · similarity 0.96)
- From Show to Tell: A Survey on Deep Learning-Based Image Captioning (TPAMI 2022 · similarity 0.88)
- Learning Visual Representations with Caption Annotations (ECCV 2020 · similarity 0.88)
- Compressing Visual-Linguistic Model via Knowledge Distillation (ICCV 2021 · similarity 0.87)
- Robust Visual Reasoning via Language Guided Neural Module Networks (NeurIPS 2021 · similarity 0.87)
- Image Captioning and Visual Question Answering Based on Attributes and External Knowledge (TPAMI 2017 · similarity 0.87)
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.
