YOLOv3 Inference Framework in C++

My journey to object detection began with YOLOv3. I think it’s really a good starting point for someone like me without computer vison background to understand what’s going on behind the scence. While learning object detection, I made a simple modified C++ version (with LibTorch) based on others’ work ^{1 2 3}.

See code here

Performance

test code:

1

yolov3 models/yolov3.cfg models/yolov3.weights images

Results:

1. tested with CPU: Core i9

   • Windows
   • MAC: average time (682 ms/image).

2. tested with GPU: Tesla V100

   • Linux: average time (22 ms/image).

Features

Supports

• NMS
• Soft NMS
• Weighted NMS

TODO

• Support training
• …

Credits

YOLOv3_CPP repo is created based on the implementations below: