InSentiCap_model

A pytorch implementation of our paper Image Captioning with Inherent Sentiment (ICME 2021 Oral).

Citation

@inproceedings{li2021image,
  title={Image Captioning with Inherent Sentiment},
  author={Li, Tong and Hu, Yunhui and Wu, Xinxiao},
  booktitle={2021 IEEE International Conference on Multimedia and Expo (ICME)},
  year={2021},
  organization={IEEE}
}

Environment

• Python 3.7
• Pytorch 1.3.1

Method

1. Architecture

2. Train Strategy

• Pre-training stage
  
• Fine-tuning stage
  

Result

Evaluation metrics

Sentiment	Bleu-1	Bleu-3	METEOR	CIDEr	ppl(↓)	cls(%)
positive	59.7	25.3	20.9	61.3	13.0	98.5
negative	59.1	24.3	19.4	53.3	12.3	95.5
neutral	73.5	41.2	24.7	97.5	8.4	98.9