euske/fastdet

Fast Object Detector (hopefully)

FastDet

Fast object detector with distributed neural network.

Prerequisites

Client

• Unity
• Barracuda (> 2.0.0)

Server

• Python

• Pillow

• ONNX Runtime (w/ GPU)

  $ pip install -r requirements.txt

Building Client (Unity)

1. Launch Unity Hub and open a project.
2. Select the "fastdet-test" folder.
3. "File" → "Open Scene" and select the "SampleScene.unity".
4. Open "Project" → "Assets" tab and make sure the "Yolov3-tiny" model is visible.
5. Select "SampleScene" → "Canvas" and make sure the Yolo Model is associated with yolov3-tiny. (if missing, click it and connect to the yolov3-tiny.onnx)
6. Connect the PC to a camera, press the Play button at the top.
7. "File" → "Build Settings" and select "Android". Press "Switch Platform".
8. Enable the "Developer Mode" and "USB Debugging" on an Android phone.
9. Press "Build & Run".

Testing

Test detector only

$ python server/detector.py -c 80 models/yolov3-full.onnx testdata/dog.jpg
$ python server/detector.py -c 9 models/yolov3-rsu.onnx testdata/rsu1.jpg

Test server with dummy detector

$ python server/server.py -s 10000
$ python server/client.py rtsp://localhost:10000/detect testdata/dog.jpg

Test server with full detector

$ python server/server.py -s 10000 full:80:models/yolov3-full.onnx rsu:9:models/yolov3-rsu.onnx
$ python server/client.py rtsp://localhost:10000/full testdata/dog.jpg
$ python server/client.py rtsp://localhost:10000/rsu testdata/rsu1.jpg

Test server w/ CUDA

$ python server/server.py -s 10000 -m cuda full:80:models/yolov3-full.onnx

Debugging on Android

> cd \Program Files\Unity\Hub\Editor\*\Editor\Data\PlaybackEngines\AndroidPlayer\SDK\platform-tools
> adb logcat -c
> adb logcat -s Unity

Running

1. launch the server.
2. open the SampleScene.unity.
3. configure the Server Url with the appropriate host/port.
4. play the scene.

TODOs

• IPv6 support (both client and server).
• Dockerize the server.
• Rewrite the server in a faster language (Go or C# maybe?).