Semi Global Matching

The Semi Global Matching (SGMDisparity) codelet is part of the Isaac AMR Navigation Stack. Its purpose is to compute the disparity output for a pair of left and right images from a HAWK camera. Improvments include:

1. Wider field of view

2. 30 FPS execution on OFA engine

3. Parameter tuning

• left_image_receiver: The receiver containing the left image data in NV12 format. Images should be 1920x1200 in size on the ORIN/XAVIER backend.

• right_image_receiver: The receiver containing the right image data in NV12 format. Images should be 1920x1200 in size on the ORIN/XAVIER backend.

• output_transmitter: The disparity output generated by the codelet

• backend: The backend that SGMDisparity will use. When running on x86, use “CUDA”; on Orin devices, use “ORIN”; and for Xavier, use “XAVIER”.

• pool: The memory allocator for the disparity output

You can use optional parameters to tune SGM performance:

• confidence_threshold: The confidence threshold for the VPI SGM algorithm. The default value is 65023.

• window_size: The window size for SGM disparity calculation. The default value is 7.

• num_passes: The number of passes that the VPI algorithm takes to compute the SGM output. The default value is 2.

• p1: The penalty on disparity changes of +/- 1 between neighboring pixels. The default value is 8.

• p2: The penalty on disparity changes of more than 1 between neighboring pixels. The default value is 120.

• p2_alpha: The alpha for P2. The default value is 1.

• quality: The quality of the disparity output. The default value is 1.

• output_transmitter: The disparity output generated by the codelet

The internal SGMDisparity algorithm consists of two steps:

1. Downsample images to 1280 x 800.

2. Use the VPI SGM algorithm to compute the disparity output