/UniRef

[ICCV2023] Segment Every Reference Object in Spatial and Temporal Spaces

Primary LanguagePythonMIT LicenseMIT

UniRef++: Segment Every Reference Object in Spatial and Temporal Spaces

Official implementation of UniRef++, an extended version of ICCV2023 UniRef.

UniRef

Highlights

  • UniRef/UniRef++ is a unified model for four object segmentation tasks, namely referring image segmentation (RIS), few-shot segmentation (FSS), referring video object segmentation (RVOS) and video object segmentation (VOS).
  • At the core of UniRef++ is the UniFusion module for injecting various reference information into network. And we implement it using flash attention with high efficiency.
  • UniFusion could play as the plug-in component for foundation models like SAM.

Schedule

  • Add Training Guide
  • Add Evaluation Guide
  • Add Data Preparation
  • Release Model Checkpoints
  • Release Code

Results

video_demo.mp4

Referring Image Segmentation

RIS

Referring Video Object Segmentation

RVOS

Video Object Segmentation

VOS

Zero-shot Video Segmentation & Few-shot Image Segmentation

zero-few-shot

Model Zoo

Objects365 Pretraining

Model Checkpoint
R50 model
Swin-L model

Imge-joint Training

Model RefCOCO FSS-1000 Checkpoint
R50 76.3 85.2 model
Swin-L 79.9 87.7 model

Video-joint Training

The results are reported on the validation set.

Model RefCOCO FSS-1000 Ref-Youtube-VOS Ref-DAVIS17 Youtube-VOS18 DAVIS17 LVOS Checkpoint
UniRef++-R50 75.6 79.1 61.5 63.5 81.9 81.5 60.1 model
UniRef++-Swin-L 79.1 85.4 66.9 67.2 83.2 83.9 67.2 model

Installation

See INSTALL.md

Getting Started

Please see DATA.md for data preparation.

Please see EVAL.md for evaluation.

Please see TRAIN.md for training.

Citation

If you find this project useful in your research, please consider cite:

@article{wu2023uniref++,
  title={UniRef++: Segment Every Reference Object in Spatial and Temporal Spaces},
  author={Wu, Jiannan and Jiang, Yi and Yan, Bin and Lu, Huchuan and Yuan, Zehuan and Luo, Ping},
  journal={arXiv preprint arXiv:2312.15715},
  year={2023}
}
@inproceedings{wu2023uniref,
  title={Segment Every Reference Object in Spatial and Temporal Spaces},
  author={Wu, Jiannan and Jiang, Yi and Yan, Bin and Lu, Huchuan and Yuan, Zehuan and Luo, Ping},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={2538--2550},
  year={2023}
}

Acknowledgement

The project is based on UNINEXT codebase. We also refer to the repositories Detectron2, Deformable DETR, STCN, SAM. Thanks for their awsome works!