Skip to content

Latest commit

 

History

History
executable file
·
85 lines (62 loc) · 5.23 KB

README.md

File metadata and controls

executable file
·
85 lines (62 loc) · 5.23 KB
Raw

GLNet for Memory-Efficient Segmentation of Ultra-High Resolution Images

Language grade: Python License: MIT

Collaborative Global-Local Networks for Memory-Efficient Segmentation of Ultra-High Resolution Images

Wuyang Chen*, Ziyu Jiang*, Zhangyang Wang, Kexin Cui, and Xiaoning Qian

In CVPR 2019 (Oral). [Youtube]

Overview

Segmentation of ultra-high resolution images is increasingly demanded in a wide range of applications (e.g. urban planning), yet poses significant challenges for algorithm efficiency, in particular considering the (GPU) memory limits.

We propose collaborative Global-Local Networks (GLNet) to effectively preserve both global and local information in a highly memory-efficient manner.

  • Memory-efficient: training w. only one 1080Ti and inference w. less than 2GB GPU memory, for ultra-high resolution images of up to 30M pixels.

  • High-quality: GLNet outperforms existing segmentation models on ultra-high resolution images.

Acc_vs_Mem
Inference memory v.s. mIoU on the DeepGlobe dataset.
GLNet (red dots) integrates both global and local information in a compact way, contributing to a well-balanced trade-off between accuracy and memory usage.

Examples
Ultra-high resolution Datasets: DeepGlobe, ISIC, Inria Aerial

Methods

GLNet
GLNet: the global and local branch takes downsampled and cropped images, respectively. Deep feature map sharing and feature map regularization enforce our global-local collaboration. The final segmentation is generated by aggregating high-level feature maps from two branches.

GLNet
Deep feature map sharing: at each layer, feature maps with global context and ones with local fine structures are bidirectionally brought together, contributing to a complete patch-based deep global-local collaboration.

Training

Current this code base works for Python version >= 3.5.

Please install the dependencies: pip install -r requirements.txt

First, you could register and download the Deep Globe "Land Cover Classification" dataset here: https://competitions.codalab.org/competitions/18468

Then please sequentially finish the following steps:

  1. ./train_deep_globe_global.sh
  2. ./train_deep_globe_global2local.sh
  3. ./train_deep_globe_local2global.sh

The above jobs complete the following tasks:

  • create folder "saved_models" and "runs" to store the model checkpoints and logging files (you could configure the bash scrips to use your own paths).
  • step 1 and 2 prepare the trained models for step 2 and 3, respectively. You could use your own names to save the model checkpoints, but this requires to update values of the flag path_g and path_g2l.

Evaluation

  1. Please download the pre-trained models for the Deep Globe dataset and put them into folder "saved_models":
  1. Download (see above "Training" section) and prepare the Deep Globe dataset according to the train.txt and crossvali.txt: put the image and label files into folder "train" and folder "crossvali"
  2. Run script ./eval_deep_globe.sh

Citation

If you use this code for your research, please cite our paper.

@inproceedings{chen2019GLNET,
  title={Collaborative Global-Local Networks for Memory-Efficient Segmentation of Ultra-High Resolution Images},
  author={Chen, Wuyang and Jiang, Ziyu and Wang, Zhangyang and Cui, Kexin and Qian, Xiaoning},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

Acknowledgement

We thank Prof. Andrew Jiang and Junru Wu for helping experiments.