I have the following task: to select all the individual items in the RGB image into separate segments, and among these items there may be items that have not been used before. I have attached an example RGB and Depth image of the scene.
I have conducted my own review of existing articles in this area. This article shows very good results: https://arxiv.org/pdf/2007.15157.pdf. In it, the authors used a pre-trained neural network as a feature extractor (since it already has weights configured to highlight image properties). Then, using this neural network, they obtained a feature map, and then clustered the resulting image and thus allocated segments.
The question is: what training data can be used to train this neural network? And how to get a single featuremap, because the neural network has several feature maps stored in filters. Unfortunately, the authors of this article do not respond to me by Email, perhaps the community will have guesses?