Research map / Vision-language and multimodal / Human and embodied modeling
Show and Tell: Lessons Learned from the 2015 MSCOCO Image Captioning Challenge
Oriol Vinyals, Alexander Toshev, Samy Bengio A, Dumitru Erhan
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 |
|---|---|---|---|
| tune and refine the system for the competition (NICv2) | NIC · MSCOCO dev | CIDEr: 85.5 -> 99.8 | Improved |
| replace GoogLeNet with Batch Normalization image model (ablation) | NIC · MSCOCO | BLEU-4: 2 | Improved |
| reduce beam size from 20 to 3 (ablation) | NIC · MSCOCO | BLEU-4: 2 | Improved |
| reduce beam size (shallower search) (ablation) | NIC · MSCOCO | repeated training captions: 80% -> 60% | Improved |
| fine-tune the image model after the LSTM has settled (ablation) | NIC · MSCOCO | BLEU-4: 1 | Improved |
| jointly train CNN and LSTM from the start (ablation) | NIC · MSCOCO | caption quality | Hurt |
| add scheduled sampling (ablation) | NIC · MSCOCO | BLEU-4: 1.5 | Improved |
| ensemble 5 scheduled-sampling and 10 fine-tuned models (ablation) | NIC · MSCOCO | BLEU-4: 1.5 | Improved |
| transfer from Flickr30k instead of MSCOCO to PASCAL (ablation) | NIC · PASCAL | BLEU-1: 59 -> 53 | Hurt |
| run MSCOCO-trained model on SBU instead of SBU-trained (ablation) | NIC · SBU | BLEU-1: 28 -> 16 | Hurt |
| train on Flickr30k (4x more data) instead of Flickr8k (ablation) | NIC · Flickr8k | BLEU: 4 BLEU points | Improved |
| use MNLM ranking-aware loss versus NIC on Flickr30k annotation | NIC · Flickr30k ranking | Image annotation R@1: 23 -> 17 | Hurt |
Closest papers by meaning
- From Show to Tell: A Survey on Deep Learning-Based Image Captioning (TPAMI 2022 · similarity 0.91)
- An Empirical Study of Language CNN for Image Captioning (ICCV 2017 · similarity 0.91)
- Convolutional Image Captioning (CVPR 2018 · similarity 0.90)
- Unsupervised Image Captioning (CVPR 2019 · similarity 0.89)
- Partially-Supervised Image Captioning (NeurIPS 2018 · similarity 0.89)
- Unified Generative and Discriminative Training for Multi-modal Large Language Models (NeurIPS 2024 · 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.
