Segmentation dataset and model for garden like scenario

[Carlos Vicente]

I've very recently started looking Deep learning, image classification/object detection and Semantic segmentation, so in general I feel I still lack a lot of bases in this overall subject. 

I've followed the Hello AI World and the 2 Days to a Demo tutorials from NVIDIA which lead to start using DIGITS.

As a "small" project I wanted to:

• build a rover to navigate autonomously around my garden (segmentation)
• chase the cat around (object detection)

For the object detection and following it around I don't have as many doubts as for the segmentation part, since I end up by building the jetbot and grasped the concept.

Nevertheless I do have a lot of doubts concerning how will I manage to train a suitable segmentation network to navigate around the garden. I thought of the following:

• Using the FOREST dataset. Although this would probably fail to its purpose because it will mark the lawn as vegetation as many of other things, so not sure how navigation would work out on a garden scenario.
• The other hypothesis would be to use a pre-train FCN model with the FOREST dataset and use a custom dataset for my garden, although, I have no idea how get the Ground-truth like images associated to the real images of my garden. Any ideas?
• Use Garden like dataset to train a network. Is there such a thing? Could anyone recommend one?
• And which pre-trained model should I use.

A more generic doubt that I also have is if I could somehow reduce the number categories in a Segmentation dataset? My first guess would be to say "no", since the Ground-truth is already given and there's nothing we can do about it (unless... we change the ground-truth images).

Any insight on any of the doubts I'm sharing here would be greatly appreciated.