Research map / Vision-language and multimodal / Human and embodied modeling
Learning Descriptive Image Captioning via Semipermeable Maximum Likelihood Estimation
Zihao Yue, Anwen Hu, Liang Zhang, Qin Jin
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 |
|---|---|---|---|
| further train BLIP with SMILE instead of MLE | BLIP · MSCOCO test | caption length: 10.0 -> 22.3 | Improved |
| further train BLIP with SMILE | BLIP · MSCOCO test | self-retrieval R@1: 6.7 -> 10.0 | Improved |
| further train BLIP with SMILE | BLIP · MSCOCO test | CLIPScore: 77.2 -> 75.0 | Hurt |
| further train BLIP with SMILE | BLIP · Flickr30K test | self-retrieval R@1: 25.4 -> 31.2 | Improved |
| train from scratch decoder with SMILE (no pre-training) (ablation) | BLIP F.S. · MSCOCO test | self-retrieval R@1: 1.9 -> 0.7 | Hurt |
| apply SMILE to pre-trained-only BLIP (ablation) | BLIP PT · MSCOCO test | self-retrieval R@1: 2.7 -> 8.5 | Improved |
| add extra MLE training steps instead of SMILE (ablation) | BLIP · MSCOCO test | self-retrieval R@1: 6.7 -> 6.5 | No clear effect |
| add initial context restriction (first-token MLE) to SMILE (ablation) | BLIP-SMILE · MSCOCO test | self-retrieval R@1: 6.8 -> 10.0 | Improved |
| first-token shifting instead of first-token MLE (ablation) | BLIP-SMILE · MSCOCO test | self-retrieval R@1: 10.0 -> 9.6 | Hurt |
| mix MLE into SMILE objective with lambda=0.01 (ablation) | BLIP-SMILE · MSCOCO test | self-retrieval R@5: 24.5 -> 25.1 | Improved |
| mix MLE into SMILE objective with lambda=0.01 (ablation) | BLIP-SMILE · MSCOCO test | CLIPScore: 75.0 -> 76.2 | Improved |
| use SMILE on MSR-VTT video captioning | BLIP4video · MSR-VTT | self-retrieval R@1: 19.0 -> 27.3 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BLIP — BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation (ICML 2022) · better on 6 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Guiding Image Captioning Models Toward More Specific Captions (ICCV 2023 · similarity 0.89)
- Uncertainty-Aware Image Captioning (AAAI 2023 · similarity 0.88)
- Contrastive Learning for Image Captioning (NeurIPS 2017 · similarity 0.88)
- Noise-Aware Image Captioning with Progressively Exploring Mismatched Words (AAAI 2024 · similarity 0.87)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.86)
- CaptionQA: Is Your Caption as Useful as the Image Itself? (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.
