Object detection

Overview

Name	Implementation	Platforms	Model	ML engine	Backend	Features
example_detection_mobilenet_ssd_v2_tflite.cpp	C++	i.MX 8M Plus i.MX 93	mobilenet_ssd_v2	TFLite	NPU (default) GPU CPU	camera gst-launch
example_detection_mobilenet_ssd_v2_tflite.sh	Bash	i.MX 8M Plus i.MX 93 i.MX 95	mobilenet_ssd_v2	TFLite	NPU (default) GPU CPU	camera gst-launch
example_detection_yolo_v4_tiny_tflite.sh	Bash	i.MX 8M Plus i.MX 93	yolov4_tiny	TFLite	NPU (default) CPU	camera gst-launch custom python tensor_filter

NOTES:

• Warmup time for NPU inference on i.MX 95 can take up to 1 minute
• No GPU support on i.MX 93
• Yolov4-tiny output does not directly work with the Yolov5 mode of tensor_decoder element, so a python filter is used to post-process and reshape this output as required

Execution

Example script can be called from target console with no further restriction. For examples that support multiple backends, default value can be overriden by explicitly defining BACKEND variable, for instance:

Bash

$ BACKEND=CPU ./detection/example_detection_mobilenet_ssd_v2_tflite.sh

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 object detection demo inference on three different hardwares:
Inference on NPU with the following script:

$ ./build/detection/example_detection_mobilenet_ssd_v2_tflite -p  ${MOBILENETV2_QUANT} -l ${COCO_LABELS} -x ${MOBILENETV2_BOXES}

For i.MX 93 use vela converted model:

$ ./build/detection/example_detection_mobilenet_ssd_v2_tflite -p  ${MOBILENETV2_QUANT_VELA} -l ${COCO_LABELS} -x ${MOBILENETV2_BOXES}

Inference on CPU with the following script:

$ ./build/detection/example_detection_mobilenet_ssd_v2_tflite -p  ${MOBILENETV2_QUANT} -l ${COCO_LABELS} -x ${MOBILENETV2_BOXES} -b CPU

Quantized model is used for better inference performances on CPU.
NOTE: inferences on i.MX8MPlus GPU have low performances, but are possible with the following script:

$ ./build/detection/example_detection_mobilenet_ssd_v2_tflite -p  ${MOBILENETV2} -l ${COCO_LABELS} -x ${MOBILENETV2_BOXES} -b GPU -n centeredReduced

The following execution parameters are available (Run ./example_detection_mobilenet_ssd_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
-l, --labels_path	Use the selected labels path
-x, --boxes_path	Use the selected boxes 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.