Awesome
Mining Latent Classes for Few-shot Segmentation
This codebase contains baseline of our paper Mining Latent Classes for Few-shot Segmentation, ICCV 2021 Oral.
Several key modifications to the simple yet effective metric learning framework:
- Remove the final residual stage in ResNet for stronger generalization
- Remove the final ReLU for feature matching
- Freeze all the BatchNorms from ImageNet pretrained model
Environment
This codebase was tested with the following environment configurations.
- Ubuntu 18.04
- CUDA 11.2
- Python 3.7.4
- PyTorch 1.6.0
- Pillow, numpy, torchvision, tqdm
- Two NVIDIA V100 GPUs
Getting Started
Data Preparation
Pretrained model: ResNet-50 | ResNet-101
Dataset: Pascal JPEGImages | SegmentationClass | ImageSets
File Organization
├── ./pretrained
├── resnet50.pth
└── resnet101.pth
├── [Your Pascal Path]
├── JPEGImages
│ ├── 2007_000032.jpg
│ └── ...
│
├── SegmentationClass
│ ├── 2007_000032.png
│ └── ...
│
└── ImageSets
├── train.txt
└── val.txt
Run the Code
CUDA_VISIBLE_DEVICES=0,1 python -W ignore main.py \
--dataset pascal --data-root [Your Pascal Path] \
--backbone resnet50 --fold 0 --shot 1
You may change the backbone
from resnet50
to resnet101
,
change the fold
from 0
to 1/2/3
, or change the shot
from 1
to 5
for other settings.
Performance and Trained Models
Here we report the performance of our modified baseline on Pascal. You can click on the numbers to download corresponding trained models.
The training time is measured on two V100 GPUs. Compared with other works, our method is efficient to train.
Setting | Backbone | Training time / fold | Fold 0 | Fold 1 | Fold 2 | Fold 3 | Mean |
---|---|---|---|---|---|---|---|
1-shot | ResNet-50 | 40 minutes | 54.9 | 66.5 | 61.7 | 48.3 | 57.9 |
1-shot | ResNet-101 | 1.1 hours | 57.2 | 68.5 | 61.3 | 53.3 | 60.1 |
5-shot | ResNet-50 | 2.3 hours | 61.6 | 70.3 | 70.5 | 56.4 | 64.7 |
5-shot | ResNet-101 | 3.5 hours | 64.2 | 74.0 | 71.5 | 61.3 | 67.8 |
Acknowledgement
We thank PANet, PPNet, PFENet and other FSS works for their great contributions.
Citation
If you find this project useful for your research, please consider citing:
@inproceedings{yang2021mining,
title={Mining Latent Classes for Few-shot Segmentation},
author={Yang, Lihe and Zhuo, Wei and Qi, Lei and Shi, Yinghuan and Gao, Yang},
booktitle={ICCV},
year={2021}
}