Research map / Vision-language and multimodal / Human and embodied modeling
Improved Image Captioning via Policy Gradient Optimization of SPIDEr
Siqi Liu, Zhenhai Zhu, Ning Ye, Sergio Guadarrama, Kevin Murphy
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 |
|---|---|---|---|
| optimize BCMR reward with policy gradient instead of MLE | Show and Tell · COCO C-5 test (leaderboard) | CIDEr-D: 0.947 -> 1.013 | Improved |
| optimize CIDEr alone with policy gradient | Show and Tell · COCO C-5 test (leaderboard) | CIDEr-D: 0.947 -> 0.995 | Improved |
| optimize BLEU-4 alone with policy gradient | Show and Tell · COCO C-5 test (leaderboard) | BLEU-4: 0.294 -> 0.346 | Improved |
| optimize BLEU-4 alone with policy gradient | Show and Tell · COCO C-5 test (leaderboard) | METEOR: 0.251 -> 0.249 | Hurt |
| use Monte Carlo rollout policy gradient instead of MIXER (SGD) | Show and Tell · COCO C-5 test (leaderboard) | CIDEr-D: 0.924 -> 1.013 | Improved |
| train MIXER with Adam instead of vanilla SGD (ablation) | MIXER-BCMR · COCO C-5 test (leaderboard) | CIDEr-D: 0.924 -> 0.991 | Improved |
| use rollout PG instead of MIXER with Adam | Show and Tell · COCO C-5 test (leaderboard) | METEOR: 0.258 -> 0.257 | No clear effect |
| optimize SPIDEr instead of BCMR | Show and Tell · COCO C-5 test (leaderboard) | CIDEr-D: 1.013 -> 1.000 | Hurt |
| optimize SPIDEr instead of BCMR | Show and Tell · Human evaluation (492 COCO test images) | fraction of not-bad captions: 4% margin | Improved |
| train with policy gradient (BCMR or SPIDEr) instead of MLE | Show and Tell · Human evaluation (492 COCO test images) | fraction of not-bad captions | Improved |
| add CIDEr to SPICE reward (SPIDEr) instead of SPICE alone (ablation) | Show and Tell · COCO hold-out validation captions | caption quality (qualitative) | Improved |
| optimize BLEU-4 with MIXER instead of MLE (original MIXER split) | MIXER · COCO 5000 validation images | BLEU-4: 0.278 -> 0.292 | Improved |
Closest papers by meaning
- Guiding Image Captioning Models Toward More Specific Captions (ICCV 2023 · similarity 0.89)
- On Diversity in Image Captioning: Metrics and Methods (TPAMI 2020 · similarity 0.89)
- Towards Unique and Informative Captioning of Images (ECCV 2020 · similarity 0.89)
- Image Captioning With Controllable and Adaptive Length Levels (TPAMI 2023 · similarity 0.89)
- BalCapRL : A Balanced Framework for RL-Based MLLM Image Captioning (NeurIPS 2026 · similarity 0.88)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · 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.
