About training the my own datasets

[xfby2016]

Excellent work！ I tried to train my own dataset (rock slice) with only two classes: background and target ，But the segmentation result of the test is very unsatisfactory，What might be the reason for this？

[jbardelli]

This proyect is a work in progress, so I'm still learning. From my short experience, the segmentation of minerals is a very difficult task compared with segmenting cars, bikes, etc. What is your backgound, other minerals or porosity? Without knowing your project is difficult to say, but consider the following: 1) It needs to be a difference between background and target. The more alike they are the more samples you need to feed your model in order to learn those small differences. 2) If there is some preprocessing you can apply to magnify differences it might help. 3) Generating masks for segmentation of thin sections is a painful task, but try to mark all the appearances of your target in order to enhance detection. 4) Sometimes if you have enough memory, increasing the batch size could help.

If you want to share more of your project, I might be able to give some more help. Regards, Julian

[xfby2016]

This is my training sample，Single and orthogonal polarizing images were available for each sample .I need to split each stone particle precisely ，Sadly, I only have 20 samples！ Thank you!

[lj-cug]

Any Progress? Could you share your experience at using the segmentation tool of mineral?