Inference using Mask-RCNN custom model

Mask-RCNN architecture can be used for multiple vision task. In this case it is used for Instance Segmentation.

Direct result:

• Binary raster (tree/non tree)

Derived results:

• Polygons vector layer

• Bounding Boxes vector layer

• Centroids vector layer

Input

For this example we will use this image of a silvopastoral system.

Input image features

Feature	Value	Unit
Spatial resolution	4.77	m
Width	853	px
Height	446	px

The images can be downloaded here.

Configuration

Input

• Select the appropiate Input layer from the dropdown menu.

• For Extent select Layer extent.

Output

• Select the appropiate Output directory and Output name.

Processing

• Go to the Inference tab

  • Select Mask-RCNN model from the dropdown menu

  • Select the desired Result types, the direct result type for this model is Binary

  • Press Process to start the inference process

Results

The resulting added layers depend on the selected Result types

Direct result:

Processing Result	Result Type	Result
RASTER LAYERS
Direct result	Binary
Derived result	Grayscale	NOT AVAILABLE FOR THIS MODEL
VECTOR LAYERS
Derived result	Polygons
Derived result	Bounding Boxes
Derived result	Centroids