Machine Learning

Z-Depth Map

New machine-learning-based models are available as Analysis Type options to generate depth data. ML Depth Indoor and ML Depth Outdoor are internal machine learning models, while Load Model allows an external ONNX depth extraction model to be loaded.

Select:	To:
Face Analysis	Extract depth from human face input content using a machine learning model.
Media Input	Use the content of a media input as Depth.
ML Depth Indoor	Extract depth using a machine learning model. Optimised for indoor scenes.
ML Depth Outdoor	Extract depth using a machine learning model. Optimised for outdoor scenes.
Load Model	Extract depth using any ONNX depth extraction model.

Note: The Global Analysis option has been removed and only appears when a legacy setup is loaded.

Selective

Default Selection

The ML menu is selected by default rather than the Keyers menu.

Depth

The ML Depth Indoor and ML Depth Outdoor machine learning models are available in the ML layer options. Use these models to create isolation and mattes using the depth data of an image. Three modes provide different ways to segment the Depth data.

For more information, see Depth.

Inference

Depth Output Normalization

Add the following output tokens to the JSON sidecar to automatically apply a min/max normalization to the model output.

"PostProcessing": "MinMaxNormalization", "PostProcessingParams":{"Range":[0.0, 1000.0]}

Resizing Filters

The Impulse and Bicubic resizing filters are available in the Inference node.

Some models, such as those that extract a Z-Depth map from an image, may require resizing but must not use filtering to produce accurate results. For these models, it is recommended to select Impulse as the output resizing filter.

Matte Output

An ONNX model that generates a single channel output automatically sets the ColorSpace to Matte.

Matchbox

The ML Depth Map Matchbox shader has been replaced by two Matchbox shaders: ML Depth Indoor and ML Depth Outdoor. These shaders provide the same result as their corresponding options in the Z-Depth map object of Action and Image.

The Matte Output section offers the same controls as the Depth models in the ML section of a Selective object, providing the ability to generate a matte from the Depth result.