This repository contains the code used in the paper "Image Segmentation Using Text and Image Prompts".
November 2022: CLIPSeg has been integrated into the HuggingFace Transformers library. Thank you, NielsRogge!
September 2022: We released new weights for fine-grained predictions (see below for details).
March 2022: The Paper has been accepted to CVPR 2022!
The systems allows to create segmentation models without training based on:
- An arbitrary text query
- Or an image with a mask highlighting stuff or an object.
In the Quickstart.ipynb notebook we provide the code for using a pre-trained CLIPSeg model. If you run the notebook locally, make sure you downloaded the rd64-uni.pth weights, either manually or via git lfs extension.
It can also be used interactively using MyBinder
(please note that the VM does not use a GPU, thus inference takes a few seconds).
This code base depends on pytorch, torchvision and clip (pip install git+https://github.com/openai/CLIP.git).
Additional dependencies are hidden for double blind review.
PhraseCutandPhraseCutPlus: Referring expression datasetPFEPascalWrapper: Wrapper class for PFENet's Pascal-5i implementationPascalZeroShot: Wrapper class for PascalZeroShotCOCOWrapper: Wrapper class for COCO.
CLIPDensePredT: CLIPSeg model with transformer-based decoder.ViTDensePredT: CLIPSeg model with transformer-based decoder.
For some of the datasets third party dependencies are required. Run the following commands in the third_party folder.
git clone https://github.com/cvlab-yonsei/JoEm
git clone https://github.com/Jia-Research-Lab/PFENet.git
git clone https://github.com/ChenyunWu/PhraseCutDataset.git
git clone https://github.com/juhongm999/hsnet.gitThe MIT license does not apply to these weights.
We provide three model weights, for D=64 (2x, ~4MB each) and D=16 (~1MB).
wget https://owncloud.gwdg.de/index.php/s/ioHbRzFx6th32hn/download -O weights.zip
unzip -d weights -j weights.zip
We introduced a more complex module for transforming tokens into predictions that allow for more refined predictions (in contrast to the square-like predictions of other weights). Corresponding weights are available in the weight download above called rd64-uni-refined.pth.
They can be loaded by:
model = CLIPDensePredT(version='ViT-B/16', reduce_dim=64, complex_trans_conv=True)
model.load_state_dict(torch.load('weights/rd64-uni-refined.pth'), strict=False)See below for a direct comparison of the new fine-grained weights (top) and the old weights (below).
To train use the training.py script with experiment file and experiment id parameters. E.g. python training.py phrasecut.yaml 0 will train the first phrasecut experiment which is defined by the configuration and first individual_configurations parameters. Model weights will be written in logs/.
For evaluation use score.py. E.g. python score.py phrasecut.yaml 0 0 will train the first phrasecut experiment of test_configuration and the first configuration in individual_configurations.
In order to use the dataset and model wrappers for PFENet, the PFENet repository needs to be cloned to the root folder.
git clone https://github.com/Jia-Research-Lab/PFENet.git
The source code files in this repository (excluding model weights) are released under MIT license.
@InProceedings{lueddecke22_cvpr,
author = {L\"uddecke, Timo and Ecker, Alexander},
title = {Image Segmentation Using Text and Image Prompts},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2022},
pages = {7086-7096}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent