Simpler is Better: Few-shot Semantic Segmentation with Classifier Weight Transformer. ICCV2021.

Introduction

We proposed a novel model training paradigm for few-shot semantic segmentation. Instead of meta-learning the whole, complex segmentation model, we focus on the simplest classifier part to make new-class adaptation more tractable. Also, a novel meta-learning algorithm that leverages a Classifier Weight Transformer (CWT) for adapting dynamically the classifier weights to every query sample is introduced to eliminate the impact of intra-class discripency.

Architecture

Environment

Other configurations can also work, but the results may be slightly different.

• torch==1.6.0
• numpy==1.19.1
• cv2==4.4.0
• pyyaml==5.3.1

Dataset

We follow the same rule to download and process dataset as that in https://github.com/Jia-Research-Lab/PFENet. After processing, please change the "data_root" and "train/val_list" in config files accordingly.

Pre-trained models in the first stage

For convenience, we provide the pre-trained models on base classes for each split. Download it here: https://drive.google.com/file/d/1VPBquiy4DZXu8b6qsSB6XtO5_6jTpXgo/view?usp=sharing, and change "resume_weights" to this folder.

Episodic training and inference

• The general training script

sh scripts/train.sh {data} {split} {[gpu_ids]} {layers} {shots}

• This is an example with 1-shot, ResNet-50, split-0 on PASCAL and GPU device [0].

sh scripts/train.sh pascal 0 [0] 50 1

• Inference script

sh scripts/test.sh {data} {shot} {[gpu_ids]} {layers} {split}

Contact

Please write down issues or contact me via zhihe.lu [at] surrey.ac.uk if you have any questions.

Citation

If you feel helpful of this work, please cite it.

@inproceedings{lu2021simpler,
  title={Simpler is Better: Few-shot Semantic Segmentation with Classifier Weight Transformer},
  author={Lu, Zhihe and He, Sen and Zhu, Xiatian and Zhang, Li and Song, Yi-Zhe and Xiang, Tao},
  booktitle={ICCV},
  year={2021}
}

Acknowledgments

Thanks to the code contributors. Some parts of code are borrowed from https://github.com/Jia-Research-Lab/PFENet and https://github.com/mboudiaf/RePRI-for-Few-Shot-Segmentation.