VinVL: Revisiting Visual Representations in Vision-Language Models

Updates

02/28/2021: Project page built.

Introduction

This repository is the project page for VinVL, containing necessary instructions to reproduce the results presented in the paper. We presents a detailed study of improving visual representations for vision language (VL) tasks and develops an improved object detection model to provide object-centric representations of images. Compared to the most widely used bottom-up and top-down model (code), the new model is bigger, better-designed for VL tasks, and pre-trained on much larger training corpora that combine multiple public annotated object detection datasets. Therefore, it can generate representations of a richer collection of visual objects and concepts. While previous VL research focuses mainly on improving the vision-language fusion model and leaves the object detection model improvement untouched, we show that visual features matter significantly in VL models. In our experiments we feed the visual features generated by the new object detection model into a Transformer-based VL fusion model OSCAR (code), and utilize an improved approach to pre-train the VL model and fine-tune it on a wide range of downstream VL tasks. Our results show that the new visual features significantly improve the performance across all VL tasks, creating new state-of-the-art results on seven public benchmarks.

Performance

t2i: text-to-image retrieval; i2t: image-to-text retrieval; IC: image captioning on COCO.

Leaderboard results

VinVL has achieved top-position in several VL leaderboards, including Visual Question Answering (VQA), Microsoft COOC Image Captioning, Novel Object Captioning (nocaps), and Visual Commonsense Reasoning (VCR).

Comparison with image features from bottom-up and top-down model (code).

We observe uniform improvements on seven VL tasks by replacing visual features from bottom-up and top-down model with ours. The NoCaps baseline is from VIVO, and our results are obtained by directly replacing the visual features. The baselines for rest tasks are from OSCAR, and our results are obtained by replacing the visual features and performing OSCAR+ pre-training. All models are BERT-Base size. As analyzed in Section 5.2 in the VinVL paper, the new visual features contributes 95% of the improvement.

Please see the following two figures for visual comparison.

Source code

Pretrained Faster-RCNN model and feature extraction

The pretrained X152-C4 object-attribute detection can be downloaded here. With code from our Scene Graph Benchmark Repo (to be released soon), one can extract features with following command:

python tools/test_sg_net.py --config-file sgg_configs/vgattr/vinvl_x152c4.yaml TEST.IMS_PER_BATCH 2 MODEL.WEIGHT models/vinvl/vinvl_vg_x152c4.pth MODEL.ROI_HEADS.NMS_FILTER 1 MODEL.ROI_HEADS.SCORE_THRESH 0.2 DATA_DIR "../maskrcnn-benchmark-1/datasets1" TEST.IGNORE_BOX_REGRESSION True MODEL.ATTRIBUTE_ON True TEST.OUTPUT_FEATURE True

The output feature will be encoded as base64.

Find more pretrained models in DOWNLOAD.

Pre-exacted Image Features

For ease-of-use, we make pretrained features and predictions available for all pretraining datasets and downstream tasks. Please find the instructions to download them in DOWNLOAD.

Pretraind Oscar+ models and VL downstream tasks

The code to produce all vision-language results (both pretraining and downstream task finetuning) can be found in our OSCAR repo. One can find the model zoo for vision-language tasks here.

Citations

Please consider citing this paper if you use the code:

@article{li2020oscar,
  title={Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks},
  author={Li, Xiujun and Yin, Xi and Li, Chunyuan and Hu, Xiaowei and Zhang, Pengchuan and Zhang, Lei and Wang, Lijuan and Hu, Houdong and Dong, Li and Wei, Furu and Choi, Yejin and Gao, Jianfeng},
  journal={ECCV 2020},
  year={2020}
}

@article{zhang2021vinvl,
  title={VinVL: Making Visual Representations Matter in Vision-Language Models},
  author={Zhang, Pengchuan and Li, Xiujun and Hu, Xiaowei and Yang, Jianwei and Zhang, Lei and Wang, Lijuan and Choi, Yejin and Gao, Jianfeng},
  journal={CVPR 2021},
  year={2021}
}