Explanation regarding Inference Results

[sunny52266]

Hi!

First of all, I would like to thank you for your work.

Currently, I am testing to train a model for parcel mapping purposes (not building extraction).

The input training data contains 5cm imagery of a small area and its relevant parcels digitized in shapefile format. (also tried geojson).

The vector dataset includes a class field.

following is a sample of the input dataset:

I tried training the model on 2, 10 and 50 epochs.

However, the resulting inference seems to segment the whole block instead of individual polygons. like the following:

Is there something I am missing or is it supposed to be like this?

[giswqs]

For semantic segmentation, adjacent objects in the binary masks are merged together when converting to polygons. This issue will be resolved when instance segmentation is supported.