How can I use my own dataset for Image segmentation using Tensorflow

Question

I have a huge problem using my own created dataset for image segmentation using Tensorflow. The dataset that I've build contain images like the one shown below:

The problem that I have is: How do I use my own dataset specifically for image segmentation? I've looked at the documentation on how to create datasets but all the examples either only use object detection with a single class or classify the entire image. I want to assign every pixel to a class (image segmentation) and use it to train my model.

I've also found examples of image segmentation but they all use existing datasets such as Cityscapes, ADE20k etc. How can I use my own images and data and transform it to a tensorflow dataset so that I can use it for training?

Answer 1

Just having segmented images is probably not enough. The training data for segmentation needs to be in a specific format. Have a look at the coco dataset for image segmentation. Sometimes we need to convert the dataset into that format. I'd suggest reading up a bit on how to train a mask rcnn model on your own dataset. There are a number of articles available online.

Answer 2

Here's a great video describing what a COCO dataset is, and Here's a video describing how to actually go about making it. There are a few options, but this one seems reasonable in terms of effort.