TensorFlow Lite : label_image object detection on Raspberry Pi Zero W

This is a follow-on to compiling TensorFlow Lite for the Raspberry Pi Zero W as detailed at https://github.com/cloudwiser/TensorFlowLiteRPIZero

Having got the C++ minimal example code compiled, this extends the label_image example to demonstrate object detection and bounding box extraction using the SSD Mobilenet model trained against the COCO traning set

Copy the content in the ./example, ./make and ./bin sub-directories of this repo to the equivalent paths in your local ./tensorflow/tensorflow/lite/... directory structure (after backing up the original content)

Please note that label_image has not been tested with other quantized or non-quantized models and it expects 1 input tensor and 4 output tensors...so running it with other models may either fail this test and/or cause a runtime error during model load or invocation

If this does occur, it should be a case of adjusting the input and output tensor size checks and ensuring the correct handling of TfLite float or uint8_t output tensors

Once you have installed (and amended) the code, build it via...

$ cd ~/tensorflow/tensorflow/lite/tools/make/

$ ./build_rpi_armv6_label_image.sh

and then run label_image as follows. Note: the -o 1 option writes the output image (with any above-threshold bounding boxes rendered) to an out file of the format out_<input_filename>

$ cd ~/tensorflow/tensorflow/lite/tools/make/gen/rpi_armv6/bin

$ ./label_image -i fruit.bmp -m ./models/coco_ssd_mobilenet_v1_1.0_quant_2018_06_29/detect.tflite -l ./models/coco_ssd_mobilenet_v1_1.0_quant_2018_06_29/labelmap.txt -v 1

$ ./label_image -i orange_banana_apple.bmp -m ./models/coco_ssd_mobilenet_v1_1.0_quant_2018_06_29/detect.tflite -l ./models/coco_ssd_mobilenet_v1_1.0_quant_2018_06_29/labelmap.txt -o 1

Typical inference speeds for the sample orange_banana_apple.bmp test image provided is ~7000 ms on a Raspberry Pi Zero W

Run label_image with -h and/or read the source code for the complete set of command line options

This application is based on thelabel_image example code in the Tensorflow Lite repo at https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/examples/label_image

Sample BMP output showing MS-COCO objects detected above the set threshold