This is Consensus-Aware Visual-Semantic Embedding (CVSE), the official source code for the paper Consensus-Aware Visual-Semantic Embedding for Image-Text Matching (ECCV 2020). It is built on top of the VSE++ in PyTorch.
Abstract:
Image-text matching plays a central role in bridging vision and language. Most existing approaches only rely on the image-text instance pair to learn their representations, thereby exploiting their matching relationships and making the corresponding alignments. Such approaches just exploit the superficial associations contained with the instance pairwise data, with no consideration of any external commonsense knowledge, which may hinder their capabilities to reason the higher-level relationships between image and text. In this paper, we propose a Consensus-Aware Visual-Semantic Embedding (CVSE) model to incorporate the consensus information, namely the commonsense knowledge shared between both modalities, into image-text matching. Specifically, the consensus information is exploited by computing statistical co-occurrence correlations between the semantic concepts from the image captioning corpus and deploying the constructed concept correlation graph to yield the consensus-aware concept (CAC) representations. Afterwards, CVSE learns the associations and alignments between image and text based on the exploited consensus as well as the instance-level representations for both modalities. Extensive experiments conducted on two public datasets verify that the exploited consensus makes significant contributions to constructing more meaningful visual-semantic embeddings, with the superior performances over the state-of-the-art approaches on the bidirectional image and text retrieval task.
The framework of CVSE:
The results on MSCOCO and Flicke30K dataset: (The inference/evaluation code has been corrected for fair comparision)
| Image-to-Text | Text-to-Image | ||||||
| Dataset | R@1 | R@5 | R@10 | R@1 | R@5 | R@10 | mR |
| MSCOCO | 74.8 | 95.1 | 98.3 | 59.9 | 89.4 | 95.2 | 85.5 |
| Flickr30k | 73.5 | 92.1 | 95.8 | 52.9 | 80.4 | 87.8 | 80.4 |
We recommended the following dependencies.
- Python 3.6
- PyTorch 1.1.0
- NumPy (>1.12.1)
- TensorBoard
- torchtext
- pycocotools
Download the dataset files. We use the image feature created by SCAN, downloaded here. All the data needed for reproducing the experiments in the paper, including image features and vocabularies, can be downloaded from:
wget https://iudata.blob.core.windows.net/scan/data.zip
wget https://iudata.blob.core.windows.net/scan/vocab.zipIn this implementation, we refer to the path of extracted files for data.zip as $data_path and files for vocab.zip to ./vocab_path directory.
- Train MSCOCO models:
Run
train_coco.py:
python train_coco.py --data_path "$DATA_PATH"- Train Flickr30K models:
Run
train_f30k.py:
python train_f30k.py --data_path "$DATA_PATH"-
Test on MSCOCO dataset:
- Test on MSCOCO 1K test set
python evaluate.py --data_path "$DATA_PATH" --data_name 'coco_precomp' --model_path './runs/coco/CVSE_COCO/model_best.pth.tar' --data_name_vocab coco_precomp --split test
- Test on MSCOCO 5K test set
python evaluate.py --data_path "$DATA_PATH" --data_name 'coco_precomp' --model_path './runs/coco/CVSE_COCO/model_best.pth.tar' --data_name_vocab coco_precomp --split testall
-
Test on Flickr30K dataset:
python evaluate.py --data_path "$DATA_PATH" --data_name 'f30k_precomp' --model_path './runs/f30k/CVSE_f30k/model_best.pth.tar' --data_name_vocab f30k_precomp --split test - Test on dataset transfer (MSCOCO-to-Flickr30K):
python evaluate.py --data_path "$DATA_PATH" --data_name 'f30k_precomp' --model_path './runs/coco/CVSE_COCO/model_best.pth.tar' --data_name_vocab coco_precomp --transfer_test --concept_path 'data/coco_to_f30k_annotations/Concept_annotations/'If CVSE is useful for your research, please cite our paper:
@article{Wang2020CVSE,
title={Consensus-Aware Visual-Semantic Embedding for Image-Text Matching},
author={Wang, Haoran and Zhang, Ying and Ji, Zhong and Pang, Yanwei and Ma, Lin},
booktitle={ECCV},
year={2020}
}
Part of our code is borrowed from SCAN. We thank to the authors for releasing codes.