yolov5-tflite-cpp

this project is an object detection app using tensorflow lite framework and yolov5 model. first you need to convert the yolov5 model from torch to tensorflow lite using the tool that comes in the yolov5 default repo.

yolov5 repo https://github.com/ultralytics/yolov5

python export.py --include tflite --int --weights [path/to/torch/model] --data [path/to/model/data]

after converting the model.pt you will end up with model.tflite.

Build

opencv and tensorflow lite are needed.

the app got tested with opencv 4.5 and tensorflow lite 2.7

tensorflow lite libs

get the libs folder from https://github.com/muhammedakyuzlu/tensorflow_lite_libs_cpp.git

you need to change the direcktory in the CMakeLists.txt 18.line

then

mkdir build && cd build \ 
cmake .. && make 

then you will have yolov5-tfLite executable file

Example

./yolov5-tfLite [path/to/model] [path/to/classes/names] [path/to/the/video/input] [path/to/the/video/output]

./yolov5-tfLite ../models/model.tflite ../models/coco.names ../media/walkingPeople.mp4 ../media/out.avi 

the output should be like this

Model loaded successfully.

Memory allocated for tensors.

Label Count: 80


Model run time 'milliseconds': 1096


Model run time 'milliseconds': 1116

.....

you will see the input video with boundring boxes around the 80 object that we trained the model over