Issue geoai-py python package: Multi-class preservation in instance_segmentation

[Rahim-Biswas]

I am training an instance segmentation model for 5 building classes using Roof_type. While training and inference run successfully, the output raster and the subsequent GeoJSON from orthogonalize lose the class values, collapsing all segmented objects into a single class.

My current workflow:

1. Export: geoai.export_geotiff_tiles(..., class_value_field='Roof_type')
2. Train: geoai.train_instance_segmentation_model(..., num_classes=5)
3. Inference: geoai.instance_segmentation(..., num_classes=5)
4. Vectorize: geoai.orthogonalize(masks_path, output_vector_path, epsilon=2)

The orthogonalize output contains values ranging from 0 to 379 instead of preserving the 5 unique Roof_type IDs (0,1,2,3,4) from the training labels. You can see in the following image the values are ranging from 0 to 379 in the output (inferred) raster, where it should have been five individual classes.

Is there a way to force instance_segmentation to output Class IDs, or a method to join the predicted class labels to the vector output?

[giswqs]

Have you tried the semantic segmentation approach? https://opengeoai.org/examples/train_segmentation_model

[Rahim-Biswas]

I haven't tried yet. I will try this approach.

[Rahim-Biswas]

Thank you! This approach is working :)

I have another question:
Is "segformer-B5" available in the geoai-py lib?