Awesome

Unsupervised-Segmentation

An implementation of Unsupervised Image Segmentation by Backpropagation - Asako Kanezaki 金崎朝子（東京大学）ICASSP. 2018.

Faster and more elegant than origin version. Speed up, 30s(origin) --> 5s(modify)

Paper: https://kanezaki.github.io/pytorch-unsupervised-segmentation/ICASSP2018_kanezaki.pdf

Original version Github: https://github.com/kanezaki/pytorch-unsupervised-segmentation

An Interpretation of this algorithm: https://zhuanlan.zhihu.com/p/68528056 (Warning: Simplified Chinese)

Requement

Necessary: Python 3, Torch 0.4

Unnecessary: skimage, opencv-python(cv2)

Getting Started

Try the high performance code written by me.

python3 demo_modify.py

class Args(object):  # You can change the input_image_path ↓
    input_image_path = 'image/woof.jpg'  # image/coral.jpg image/tiger.jpg

Or you want to try the code written by the original author.

python3 demo_origin.py 
python3 demo_origin.py --input image/woof.jpg

Run this demo, and press WASDQE on the keyboard to adjust the parameters. The image show in the GUI, and the parameters show in terminal in real time. You could choose Algorithm felz or Algorithm slic by commenting the code.

W,S --> parameter 1
A,D --> parameter 2
Q,E --> parameter 3

python3 demo_pre_seg__felz_slic.py

Preview

The iterative process: Save the result when the iter_number == 1,2,4,8,16,32,64,128.

The different result of Algorithm felz or Algorithm slic with different parameters.

The left picture: compactness = 10000

The right picture: compactness = 1000

The left picture: Algorithm slic

The right picture: Algorithm felz

Translate 翻译

If you can understand English, then I know you can understand this line of words (and you see this line on GitHub.)

如果你可以看得懂中文，那么我对这个算法的分析写在知乎上了（或者你就是从知乎过来的）

An implementation of Unsupervised Image Segmentation by Backpropagation

无监督图片语义分割，复现并魔改Github上的项目 https://zhuanlan.zhihu.com/p/68528056

In my opinion, this algorithm is well suited for unsupervised segmentation of satellite images, because satellite images have no directionality. It is suitable for this algorithm with a priori assumption. (Priori Assumptions: In general, the regions with the same semantic information on the satellite images tend to occurs in a continuous area)