Research map / Vision-language and multimodal / Human and embodied modeling
Boosting Image Captioning With Attributes
Ting Yao, Yingwei Pan, Yehao Li, Zhaofan Qiu, Tao Mei
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 |
|---|---|---|---|
| feed attributes instead of image representations only | LSTM · COCO Karpathy test (GoogleNet) | CIDEr-D: 84.3 -> 96 | Improved |
| add attributes after image representations (A2 vs A1) (ablation) | LSTM-A1 · COCO Karpathy test (GoogleNet) | CIDEr-D: 96 -> 97.5 | Improved |
| feed attributes first, then image representations (A3 vs A2) (ablation) | LSTM-A2 · COCO Karpathy test (GoogleNet) | CIDEr-D: 97.5 -> 98.4 | Improved |
| input image representations at each time step (A4 vs A3) (ablation) | LSTM-A3 · COCO Karpathy test (GoogleNet) | CIDEr-D: 98.4 -> 90.6 | Hurt |
| input attributes at each time step (A5 vs A3) (ablation) | LSTM-A3 · COCO Karpathy test (GoogleNet) | CIDEr-D: 98.4 -> 98.6 | Improved |
| input attributes at each time step (A5 vs A3) (ablation) | LSTM-A3 · COCO Karpathy test (GoogleNet) | BLEU@4: 32.1 -> 32.5 | Improved |
| input image representations at each time step (A4 vs A3) (ablation) | LSTM-A3 · COCO Karpathy test (GoogleNet) | BLEU@4: 32.1 -> 30.6 | Hurt |
| add attributes as auxiliary input in LSTM-A5 vs ATT semantic attention | ATT · COCO Karpathy test (GoogleNet) | BLEU@4: 30.4 -> 32.5 | Improved |
| add attributes to LRCN-style per-step image input (A4 vs LRCN) | LRCN · COCO Karpathy test (GoogleNet) | BLEU@4: 21 -> 30.6 | Improved |
| train attribute detectors with ResNet-152 instead of GoogleNet | LSTM-A3 · COCO Karpathy test (GoogleNet) | CIDEr-D: 98.4 -> 104.9 | Improved |
| vary beam size k from 1 to 5 (ablation) | LSTM-A3 / LSTM-A5 · COCO Karpathy test (GoogleNet) | CIDEr-D | Mixed |
Closest papers by meaning
- Video Captioning With Transferred Semantic Attributes (CVPR 2017 · similarity 0.89)
- Convolutional Image Captioning (CVPR 2018 · similarity 0.89)
- Incorporating Copying Mechanism in Image Captioning for Learning Novel Objects (CVPR 2017 · similarity 0.88)
- An Empirical Study of Language CNN for Image Captioning (ICCV 2017 · similarity 0.88)
- Reference Based LSTM for Image Captioning (AAAI 2017 · similarity 0.87)
- Pointing Novel Objects in Image Captioning (CVPR 2019 · 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.
