search

asmirnou / watsor

Object detection for video surveillance
MIT License
252 stars 32 forks source link

Implementing Yolov4 for Detection #13

Closed kidhasmoxy closed 3 years ago

kidhasmoxy commented 4 years ago

Hello, Looking to implement custom models, but not clear on how that would happen when using Docker. I've built a TRT engine using https://github.com/jkjung-avt/tensorrt_demos#yolov4, but I haven't been able to translate it to Watsor.

Is it a matter of renaming the trt to gpu.buf and then dropping the libyolo_layer.so into a folder in the image?

asmirnou commented 4 years ago

Watsor uses Google models of Single-Shot Detector (SSD) represented in several formats. The original model was in protobuf and then converted to others like lite and uff to support various devices. The application expects the specific output after inference, I'm afraid other models like YOLO won't work.

The models were trained to recognize 90 classes hard-coded in the app for now. Custom models may need other classes, but there is no way to import them yet. The detection will still work, but the rendered labels won't match.

Custom models, trained based on those listed in object detection models section, can be mounted as the volume at /usr/share/watsor/model.