author is leilei
Restart this project from 2017-10-01
TODO
The latest version of the code cannot be executed and is still being updated.
python: 3.6+
ubuntu16.04 or 18.04
pytorch 1.6+ (cuda10.2 docker)
tensorboard 2.0+
- If a black border is introduced, it will be regarded as one type, and the default is 0 !
- label value is [1, N], 0 is black border class !
- Not supporting distributed(NCCL)
-
Train
-
Test
-
Dataset Details
root：
images:
labels:
train.txt：
/home/dataset/seg/images/train/aaa.jpg
/home/dataset/seg/images/train/bbb.jpg
test.txt：
/home/dataset/seg/images/test/ccc.jpg
how to match images and labels?
'/home/dataset/seg/images/train/aaa.jpg'.replace('images', 'labels')
or
'/home/dataset/seg/labels/train/aaa.jpg'.replace('.jpg', '.png')
data enhancement:
random flip, rotate, crop, noise,
hue-brightness-contrast-saturation, zoom(in out), copy-paste?, mosaic?
- TODO