Research map / Vision-language and multimodal / Human and embodied modeling
Learning to Collocate Neural Modules for Image Captioning
Xu Yang, Hanwang Zhang, 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 OBJECT, ATTRIBUTE, RELATION and FUNCTION modules instead of a single OBJECT module (ablation) | Module/O · MS-COCO Karpathy split (RL) | CIDEr-D: 123.1 -> 125.7 | Improved |
| use soft module fusion instead of hard Gumbel-Softmax selection (ablation) | Col/H · MS-COCO Karpathy split (RL) | CIDEr-D: 124.7 -> 125.7 | Improved |
| use soft module fusion instead of all-ones fusion weights (ablation) | Col/1 · MS-COCO Karpathy split (RL) | CIDEr-D: 125.3 -> 125.7 | No clear effect |
| add linguistic loss on module controller (ablation) | Col/S · MS-COCO Karpathy split (RL) | CIDEr-D: 125.7 -> 126.4 | Improved |
| add linguistic loss on module controller (ablation) | Col/H · MS-COCO Karpathy split (RL) | CIDEr-D: 124.7 -> 125.4 | Improved |
| stack decoder 3 times instead of 1 (ablation) | Col/S+L · MS-COCO Karpathy split (RL) | CIDEr-D: 126.4 -> 127.4 | Improved |
| stack the top-down LSTM 3 times without modules (ablation) | Module/O · MS-COCO Karpathy split (RL) | CIDEr-D: 123.1 -> 124.3 | Improved |
| add the four-module collocation to a 3-layer decoder (ablation) | Module/O#3 · MS-COCO Karpathy split (RL) | CIDEr-D: 124.3 -> 127.4 | Improved |
| add SGAE dictionary preserving language bias (ablation) | CNM#3 · MS-COCO Karpathy split (RL) | CIDEr-D: 127.4 -> 127.9 | Improved |
| train with only 1 sentence per image (ablation) | CNM#3 · MS-COCO Karpathy split (RL, fewer training sentences) | CIDEr-D: 127.4 -> 123.8 | Hurt |
| train with only 1 sentence per image (ablation) | Module/O · MS-COCO Karpathy split (RL, fewer training sentences) | CIDEr-D: 123.1 -> 116.3 | Hurt |
| add ATTRIBUTE and RELATION modules via full CNM instead of OBJECT module only (ablation) | Module/O · MS-COCO part-of-speech recall | noun recall (%): 44.5 -> 47.3 | Improved |
Closest papers by meaning
- Learning to Collocate Visual-Linguistic Neural Modules for Image Captioning (IJCV 2022 · 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 0.87)
- Cross-Domain Image Captioning With Discriminative Finetuning (CVPR 2023 · similarity 0.87)
- Reflective Decoding Network for Image Captioning (ICCV 2019 · similarity 0.87)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 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.
