Multiple Classes with Semantic Segmentation

[natedailey7]

I'm trying to input multiple classes into my semantic segmentation model, and I'm having trouble passing my class information into my SemanticSegmentationSlidingWindowGeoDataset. Previously, I asked the same question about ObjectDetectionSlidingWindowGeoDataset's, but I'm having trouble migrating this technique to semantic segmentation.

Here's what I'm trying:

class_config = ClassConfig(
    names=['class1', 'class2', 'class3'],
    colors=['lightgreen', 'lightblue', 'darkred'],
    )

train_image_uri = "path"
train_label_uri = "path"
val_image_uri   = "path"
val_label_uri   = "path"

class SemanticSegmentationSlidingWindowGeoDatasetCustom(SemanticSegmentationSlidingWindowGeoDataset):
	def __getitem__(self, key):
		x, y = super().__getitem__(key)
		y = y.long()
		return x, y

train_ds = SemanticSegmentationSlidingWindowGeoDatasetCustom.from_uris(
    class_config=class_config,
    image_uri=train_image_uri,
    label_vector_uri=train_label_uri,
    label_vector_source_kw=dict(
          transformers=[ClassInferenceTransformerConfig(
              class_id_to_filter={0: ['==', 'id', 0], 1: ['==', 'id', 1], 3: ['==', 'id', 3]})]),  # where 'id' is the name of the field containing my classes in the .geojson
    size=400,
    stride=400,
    transform=A.Resize(400, 400)
)

Which yields the error:

TypeError: GeoJSONVectorSource.__init__() got an unexpected keyword argument 'transformers'

After reading the documentation for SemanticSegmentationSlidingWindowGeoDataset and GeoJSONVectorSource, I'm still not sure where to place my ClassInferenceTransformerConfig. Any tips would be greatly appreciated.

[AdeelH]

Yeah, I agree that is confusing. This is because of a discrepancy between the implementations of make_od_scene and make_ss_scene. For semantic segmentation, you'll have to do it like this:

train_ds = SemanticSegmentationSlidingWindowGeoDatasetCustom.from_uris(
    ...
    label_vector_source_kw=dict(
          vector_transformers=[ClassInferenceTransformer(
              default_class_id=None,
              class_id_to_filter={0: ['==', 'id', 0], 1: ['==', 'id', 1], 3: ['==', 'id', 3]})]), 
    ...
)

• transformers --> vector_transformers
• ClassInferenceTransformerConfig --> ClassInferenceTransformer

[natedailey7]

This worked perfectly for me, thanks so much!