This YOLOv3 repo is secretly a YOLOv5 repo in disguise

What it does

Ultralytics YOLOv3 is a PyTorch implementation of the original YOLOv3 object-detection network. It handles training on standard datasets like COCO and runs inference across images, video files, webcams, and network streams. The project also provides export paths to ONNX, CoreML, and TFLite for deployment outside the training environment.

The interesting bit

The repository is explicitly built atop the YOLOv5 codebase—its own README notes the dependencies are originally from YOLOv5 and recommends tailored testing for YOLOv3. That makes it less a pure historical artifact and more a compatibility layer, letting legacy v3 weights ride on a newer, better-maintained infrastructure.

Key highlights

• Load pretrained yolov3, yolov3-spp, or yolov3-tiny weights directly through PyTorch Hub.
• Export trained models through an ONNX, CoreML, and TFLite pipeline for edge and mobile deployment.
• detect.py handles diverse input sources including screen capture, YouTube URLs, and RTSP streams.
• Training commands automatically fetch the COCO dataset and required model configurations.

Caveats

• The README states the repository is “originally YOLOv5,” and compatibility with YOLOv3 is recommended but not guaranteed.
• YOLOv3-specific documentation is limited; most tutorials and guides reference YOLOv5 examples instead.
• Ultralytics now actively develops YOLO11 in a separate repository, so this project appears to be in maintenance mode.

Verdict

Reach for this if you are stuck maintaining a legacy YOLOv3 pipeline or need v3-specific weights exported through modern converters. If you are starting a new project, the main Ultralytics repository and newer YOLO versions are the obvious choice.