train，train_annotation，test，test_annotation about image-segmentation-keras

[MjdMahasneh]

try reading the readme file (main page) section 'Preparing the data for training' they also provide a nice simple code to explain the labeling method.

[saramsv]

@duyanfang123 Did you find the answer for your question? I am having the same problem.

[lijiashu]

how can i run this code?

[saramsv]

@lijiashu
For training, predicting and visualizing, you can pretty much use the same code provided in the readme(main page). For me, I ended up using 224*224 for the image size though.
The only confusing part for me was how to provide annotation for my own data which I later figured it out. If that's the case for you, I can help you with that.

[rhs082]

saramsv:

For annotation please suggest software how for non standard data.

[saramsv]

I don't know if you can do anything without annotation. But I think you can probably use LabelMe to annotate your data.
When you have annotations for your data, then you need to create a png image corresponding to each image. The created images are going to keep your annotation information for each image.
Here is an example of how to create each annotation image.
For example you have a 224 by 224 image and it has only 2 classes (say road and grass). You create a numpy array of zeros with size 224 by 224. Then for all pixels belonging to class "road" you change the pixels' value to 1 and for all pixels belonging to class "grass" you assign value 2. The resulted image would be the annotation for the corresponding image and would have 0 for background, 1 for road and 2 for grass.

[rhs082]

Do you prefer a VGG annotator for labeling?

[lijiashu]

I had trained teh model and when i use this model to predict the picture i had not get the ringht mark.

[YLONl]

@duyanfang123 @saramsv There are many softwares for annotation, like labelme,labelImg,yolo_mark..https://cloud.tencent.com/developer/news/241350. Or use code for annotation,like the readme

[srinivas1746]

Hi @divamgupta , @saramsv , I understood, how to generate label image, but unable to give the label name(class name). Where can I give the names of the class for training own data set.

Please help me out.

[divamgupta]

You do not give the names of the labels to the model. Just make sure each segmentation class is mapped to a unique ID.

[srinivas1746]

Thanks for your quick reply.

Trained my data by giving unique pixel value for class.

For placing boundary boxes with class name, we have to see manually class color in output predicted image. its like post processing step. Am i correct?
Thanks

[divamgupta]

If you save as jpg image then you have to read the color values . Using the python interface, the predict function would return a numpy array with the class ID for each pixel .

[srinivas1746]

Okay, I will do that.

Thanks.