The code is to get segmentation image by darknet
In the process of my project, I have referenced nithi89/unet_darknet in some points and nithilan has give me many important
advices, thanks to him, and if you have interest you can visit his homepage.
This is my third version, I added dilation convolutional, and now it has not so perfect result, but i think it's good enough for me.
I will continue to update afterwards, please stay tuned.
Compile:
make -j8
Train:
./darknet segmenter train cfg/maskyolo.data cfg/instance_segment.cfg [pretrain weights file I gave to you]
Test:
./darknet segmenter test cfg/maskyolo.data cfg/instance_segment.cfg [weights file] [image path]
Merge two images:
python Merge.py
And you will get the mask image named final.png
Test image: Merge them together image:
Output image:(for orig)
Merge them together image:(not so good, 1. more epochs; 2. deeper or more complex backbone)
- https://www.dropbox.com/sh/9wrevnyzwfv8hg7/AAA1MJElri9aROsjaPTxO5KCa?dl=0
- https://pan.baidu.com/s/15gcrXGzb-fY2vGdl4KlLqg password: bk01
The Way is so easy, you only need three files:
original colorful image;
label image(pixel value is 0, 1, 2, 3 if you have 3 classes + background);
train.list.
For example with cityscape dataset:
I only have one class so the label image, as 0 is background and others are multi classes. If you have 2 classes, the label image pixel value should be 012 and so on:
1. prepare train images and label images like above images
I have added below function call in my code according to my pictures, you have to change it according to your image name
#######################################################
find_replace(labelpath, "_leftImg8bit.png", "_gtFine_labelIds.png", labelpath);
#######################################################
2. put label images and original images together
3. generate the train.list file just like:
/home/user/Desktop/YOLO_train/leftImg8bit/train/aachen_resize/jena_000012_000019_leftImg8bit.png
4. start train
./darknet segmenter train [data_file path] cfg/segment.cfg [pretrain weights file I gave to you]
If you want to see my Result Video, I have put it in: https://pan.baidu.com/s/1uJwFYLHEQ9DGFZ8RkGuagg, and the password is: ic3q
If you want to get how to change the code, see https://github.com/ArtyZe/yolo_segmentation/blob/master/Train_Details.md
What I did to change Yolo for image segmentation, I have written a blog in: https://blog.csdn.net/Artyze/article/details/82721147
After I will do some work in semantic segmentation with yolo.
If you want to do something with Yolo with me, contact me with E-mail: [email protected].