CLIP-ViL

In our paper "How Much Can CLIP Benefit Vision-and-Language Tasks?", we show the improvement of CLIP features over the traditional resnet features on the visual question answering, image captioning, navigation and visual entailment tasks.

We release the extracted features and reproducible code here.

Specifically, we develop our methods in two scenarios: (1) direct task-specific fine-tuning; and (2) Vision and Language pre-training.

CLIP-ViL-Direct/VLN

We directly plug CLIP into tasks-pecific models and finetune on three representative tasks including Visual Question Answering, Image Captioning, and Vision-Language Navigation.

Please see the corresponding code directory for full details.

Noted that in direct finetuning, for Visual Question Answering on VQA 2.0 test-dev, we are able to achieve up to 68.37% accuracy with Pythia, 74.01% accuracy with MCAN and generally more than 4.0% improvements in accuracy; For Image Captioning on Karpathy's test split of MS COCO, we got 2.1% improvements in CIDEr metric over resnet alternatives; For Navigation, On RxR, we got 5% improvements with the nDTW metric (the main metric for RxR). On R2R, we got about 6% improvements in accuracy regarding our strong baselines.

CLIP-ViL-Pretrain

In order to test the potential of combining CLIP pre-training and Vision and Language pre-training. We introduce CLIP-ViL-Pretrain, a vision-and-language model pre-trained on image-text data with CLIP visual encoder as its visual backbone. CLIP-ViL-Pretrain is pretrained on aligned image-text data with a reconstructive objective and an image-text matching objective. It is further finetuned on VQA, SNLI-VE and GQA tasks.

Please see the corresponding code directory for full details.

Noted that CLIP-ViL-Pretrain is able to achieve 76.48% accuracy on VQA 2.0 test-dev and 76.70% accuracy on test-std; 80.61% accuracy on SNLI-VE Dev and 80.20% on Test-P; 61.42% accuracy on GQA test-dev and 62.93% accuracy on test-std.

Related Links

• CLIP: paper, code
• Grid Features: paper, code

Reference

If you use CLIP-ViL in your research or wish to refer to the baseline results published here, please use the following BibTeX entry.

@misc{shen2021clip,
    title={How Much Can CLIP Benefit Vision-and-Language Tasks?}, 
    author={Sheng Shen and Liunian Harold Li and Hao Tan and Mohit Bansal and Anna Rohrbach and Kai-Wei Chang and Zhewei Yao and Kurt Keutzer},
    year={2021},
    eprint={2107.06383},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}