RegionCLIP: Region-based Language-Image Pretraining

This is the official PyTorch implementation of RegionCLIP (CVPR 2022).

Paper | Demo on Hugging Face | Slides

RegionCLIP: Region-based Language-Image Pretraining (CVPR 2022)
Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao

Overview

We propose RegionCLIP that significantly extends CLIP to learn region-level visual representations. RegionCLIP enables fine-grained alignment between image regions and textual concepts, and thus supports region-based reasoning tasks including zero-shot object detection and open-vocabulary object detection.

• Pretraining: We leverage a CLIP model to match image regions with template captions, and then pretrain our model to align these region-text pairs.
• Zero-shot inference: Once pretrained, the learned region representations support zero-shot inference for object detection.
• Transfer learning: The learned RegionCLIP model can be further fine-tuned with additional object detection annotations, allowing our model to be used for fully supervised or open-vocabulary object detection.
• Results: Our method demonstrates state-of-the-art results for zero-shot object detection and open-vocabulary object detection.

Updates

• :collision: [10/05/2022] RegionCLIP now supports not only resnet but also many vision transformers (e.g., vit, swin, davit, focalnet) for zero-shot object detection! Please checkout the zero-shot branch!
• [09/23/2022] As requested by researchers, we release the configs and scripts of pre-training. A full tutorial and pre-training data will be released later. Stay tuned!
• [09/18/2022] Organizing ECCV Workshop Computer Vision in the Wild (CVinW), where two challenges are hosted to evaluate the zero-shot, few-shot and full-shot performance of pre-trained vision models in downstream tasks:
  • Image Classification in the Wild (ICinW) Challenge evaluates on 20 image classification tasks.
  • Object Detection in the Wild (ODinW) Challenge evaluates on 35 object detection tasks.
• [07/11/2022] We included the scripts for concept feature extraction. It can be used for your own costomized concept pool!
• [07/07/2022] We included the scripts for region feature extraction. The extracted visual features can be used for various downstream tasks!
• [06/24/2022] We released a Web demo using Gradio on Hugging Face. It uses our pretrained RegionCLIP for zero-shot inference. Check it out!
• [06/20/2022] We released models and inference code for our RegionCLIP!

Outline

1. Installation
2. Datasets
3. Model Zoo
4. Zero-shot Inference
5. Transfer Learning
6. Extract Region Features
7. Extract Concept Features
8. Citation and Acknowledgement
9. Contributing

Installation

Check INSTALL.md for installation instructions.

Datasets

Check datasets/README.md for dataset preparation.

Model Zoo

Check MODEL_ZOO.md for our pretrained models.

Zero-shot Inference

After pretraining, RegionCLIP can directly support the challenging zero-shot object detection task without finetuning on detection annotation. Given an input image, our pretrained RegionCLIP can match image region features to object concept embeddings, and thus recognize image regions into many object categories. The image regions are produced by a region localizer (e.g., RPN), where the object class names come from a dictionary specifiied by users.

Visualization on custom images

We provide an example below for zero-shot object detection with pretrained RegionCLIP on custom images and for visualizing the results.

The visualized images will be placed at ./output/regions/. The visualized images would look like:

In this example, the detection results come from our pretrained RegionCLIP with ResNet50x4 architecture. The regions are proposed by an RPN trained by 866 object categories from LVIS dataset. For now, we use 1203 object class names from LVIS dataset for this visualization example. We also include an example in visualize_zeroshot_inference.sh with our pretrained RegionCLIP (ResNet50 architecture).

Evaluation for zero-shot inference

We provide an example below for evaluating our pretrained RegionCLIP (ResNet50) using ground-truth boxes on COCO dataset. This will reproduce our results in Table 4 of the paper.

For more examples, please refer to test_zeroshot_inference.sh. This script covers a wide combination of pretrained models (ResNet50, ResNet50x4), datasets (COCO, LVIS) and region proposal types (ground-truth regions, RPN proposals). Also, please refer to MODEL_ZOO.md for available trained models and datasets/README.md for setting up COCO and LVIS datasets.

Transfer Learning

Our pretrained RegionCLIP can be further fine-tuned when human annotations of objects are available. In this transfer learning setting, we demonstrate results on open-vocabulary object detection, where the object detector is trained on base categories and evaluated on both base and novel categories.

We show an example for running a trained detector on custom images. Further, we provide scripts of training and evaluation for the benchmark of open-vocabulary object detection, including COCO and LVIS datasets (Table 1 & 2 in paper).

Visualization on custom images

We provide an example below for running a trained open-vocabulary object detector on custom images and for visualizing the results. In this example, the detector is initialized using RegionCLIP (RN50x4), trained on 866 LVIS base categories, and is tasked to detect all 1203 categories on LVIS.

The visualized images will be placed at ./output/regions/.

Evaluate the trained detectors

We provide an example below for evaluating our open-vocabulary object detector, initialized by RegionCLIP (ResNet50) and trained on COCO dataset.

For more examples, please refer to test_transfer_learning.sh. This script includes benchmark evaluation for various combination of trained detectors (ResNet50, ResNet50x4) and datasets (COCO, LVIS). Also, please refer to MODEL_ZOO.md for available trained models and datasets/README.md for setting up COCO and LVIS datasets.

Train detectors on your own

We provide an example below for training an open-vocabulary object detector on COCO dataset, with pretrained RegionCLIP (ResNet50) as the initialization.

For more examples, please refer to train_transfer_learning.sh. This script provides training scripts for various combination of detector backbones (ResNet50, ResNet50x4) and datasets (COCO, LVIS). Also, please refer to MODEL_ZOO.md for available trained models and datasets/README.md for setting up COCO and LVIS datasets.

Extract Region Features

We provide scripts for extracting region features from our pre-trained RegionCLIP. Given a folder of images, our scripts extract region features (along with other detection results such as box coordinates) and save them as local files.

The following is an example using pretrained RegionCLIP with ResNet50. We extend the scripts from zero-shot inference (section above) with minor changes, such as the input and output folders.

Extract Concept Features

Along with the region feature extraction, we also provide scripts for extracting concept features from our pre-trained RegionCLIP. Given a list of concepts, our scripts extract textual embeddings and save them as local files. The following is an example using pretrained RegionCLIP. We extend the scripts from region feature extraction (section above) with minor changes.

Citation and Acknowledgement

Citation

If you find this repo useful, please consider citing our paper:

@inproceedings{zhong2022regionclip,
  title={Regionclip: Region-based language-image pretraining},
  author={Zhong, Yiwu and Yang, Jianwei and Zhang, Pengchuan and Li, Chunyuan and Codella, Noel and Li, Liunian Harold and Zhou, Luowei and Dai, Xiyang and Yuan, Lu and Li, Yin and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16793--16803},
  year={2022}
}

Acknowledgement

This repository was built on top of Detectron2, CLIP, OVR-CNN, and maskrcnn-benchmark. We thank the effort from our community.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.