depth

Depth estimation

Overview

Name	Implementation	Platforms	Model	ML engine	Backend	Features
example_depth_midas_v2_tflite.cpp	C++	i.MX 8M Plus i.MX 93	Midas v2	TFLite	NPU (default) GPU CPU	camera gst-launch custom C++ decoding

NOTES:

• No GPU support on i.MX 93
• Distances shown on screen are relative and not absolute from the camera
• The whitest, the closest

Execution

C++

C++ example script needs to be generated with cross compilation. setup_environment.sh script needs to be executed in nxp-nnstreamer-examples folder to define data paths:

$ . ./tools/setup_environment.sh

It is possible to run the depth demo inference on three different hardwares:
Inference on NPU with the following script:

$ ./build/depth/example_depth_midas_v2_tflite -p ${MIDASV2}

For i.MX 93 use vela converted model:

$ ./build/depth/example_depth_midas_v2_tflite -p ${MIDASV2_VELA}

Inference on CPU with the following script:

$ ./build/depth/example_depth_midas_v2_tflite -p ${MIDASV2} -b CPU

NOTE: Inference on i.MX8MPlus GPU is possible but not recommended because of low performances:

$ ./build/depth/example_depth_midas_v2_tflite -p ${MIDASV2} -b GPU

The following execution parameters are available (Run ./example_depth_midas_v2_tflite -h to see option details):

Option	Description
-b, --backend	Use the selected backend (CPU, GPU, NPU) default: NPU
-n, --normalization	Use the selected normalization (none, centered, reduced, centeredReduced, castInt32, castuInt8) default: none
-c, --camera_device	Use the selected camera device (/dev/video{number}) default: /dev/video0 for i.MX 93 and /dev/video3 for i.MX 8MP
-p, --model_path	Use the selected model path
-d, --display_perf	Display performances, can specify time or freq
-t, --text_color	Color of performances displayed, can choose between red, green, blue, and black (white by default)
-g, --graph_path	Path to store the result of the OpenVX graph compilation (only for i.MX8MPlus)

Press Esc or ctrl+C to stop the execution of the pipeline.