This repository contains a proof of concept application written in the Kotlin programming language for running the Yolo object detection model on a video stream. The project uses an experimental tensorflow based library to load and execute the model.
To enable GPU, compile and add to the lib directory
the Tensorflow for Java native library
(libtensorflow_jni.so and libtensorflow_framework.so).
The frozen Tensorflow computation graphs have to be available in
the yolo directory as yolo.pband tiny_yolo.pb,
which can be obtained from
this
repository.
Once Tensorflow for Java and the frozen computation graphs are set up, the application can be run using the gradle wrapper:
./gradlew run
And for tiny Yolo:
./gradlew run --args='--tiny'
By default, the first available camera device will be used
(The webcam on a laptop for example). A url to a video
can be provided with the --input argument.