https://github.com/cloudwiser/objectdetectionrpizero

TensorFlow Lite object detection example for Raspberry Pi Zero
https://github.com/cloudwiser/objectdetectionrpizero

armv6 cnn-for-visual-recognition mobilenet-ssd object-detection raspberry-pi-zero raspberry-pi-zero-w ssd-mobilenet tensorflow-examples tensorflow-lite

Last synced: 9 days ago
JSON representation

TensorFlow Lite object detection example for Raspberry Pi Zero

• Host: GitHub
• URL: https://github.com/cloudwiser/objectdetectionrpizero
• Owner: cloudwiser
• License: apache-2.0
• Created: 2019-03-21T00:15:27.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2019-11-10T18:39:57.000Z (almost 5 years ago)
• Last Synced: 2024-09-22T08:30:28.900Z (13 days ago)
• Topics: armv6, cnn-for-visual-recognition, mobilenet-ssd, object-detection, raspberry-pi-zero, raspberry-pi-zero-w, ssd-mobilenet, tensorflow-examples, tensorflow-lite
• Language: C++
• Size: 5.28 MB
• Stars: 25
• Watchers: 2
• Forks: 7
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        


### 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...

```sh

$ 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_` 

```sh

$ 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 the`label_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