Research map / Vision-language and multimodal / Human and embodied modeling
Learning to Evaluate Image Captioning
Yin Cui, Guandao Yang, Andreas Veit, Xun Huang, Serge Belongie
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 |
|---|---|---|---|
| Add pathological data augmentation during training (ablation) | Learned caption metric · COCO validation M1 | Pearson's ρ: 0.821 -> 0.939 | Improved |
| Add pathological data augmentation during training (ablation) | Learned caption metric · COCO validation M2 | Pearson's ρ: 0.807 -> 0.949 | Improved |
| Include image features in the critique context (ablation) | Learned caption metric · COCO validation | human-versus-generated caption discrimination | Improved |
| Add a reference caption to image-only context (ablation) | Learned caption metric · COCO validation | human-versus-generated caption discrimination | No clear effect |
| Train without data sampling or augmentation (ablation) | Learned caption metric · Random-caption, random-word, and word-permutation robustness tests | robustness score | Hurt |
| Use a linear classifier with concatenated context and candidate features instead of nonlinear interactions (ablation) | Learned caption metric with data augmentation · Random-caption robustness test | robustness | Hurt |
| Increase LSTM layers or hidden-feature size (ablation) | Learned caption metric · Capability evaluation | discrimination performance | No clear effect |
| Use data augmentation rather than training without it (ablation) | Learned caption metric · COCO validation M4 detailness | Pearson's ρ: 0.844 -> 0.626 | Hurt |
| Use data augmentation rather than training without it (ablation) | Learned caption metric · COCO validation M3 correctness | Pearson's ρ: 0.430 -> 0.720 | 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.
- SPICE — SPICE: Semantic Propositional Image Caption Evaluation (ECCV 2016) · better on 3 of 3 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
- Polos: Multimodal Metric Learning from Human Feedback for Image Captioning (CVPR 2024 · similarity 0.88)
- CaptionQA: Is Your Caption as Useful as the Image Itself? (CVPR 2026 · similarity 0.87)
- CAPability: A Comprehensive Visual Caption Benchmark for Evaluating Both Correctness and Thoroughness (NeurIPS 2025 · similarity 0.86)
- Painting with Words: Elevating Detailed Image Captioning with Benchmark and Alignment Learning (ICLR 2025 · similarity 0.86)
- Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating (ICLR 2025 · similarity 0.86)
- G-VEval: A Versatile Metric for Evaluating Image and Video Captions Using GPT-4o (AAAI 2025 · 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.
