What it does

This repo ports the original YOLO real-time object detector to TensorFlow. You feed it an image; it spits out bounding boxes, class labels, and confidence scores. It wraps three pretrained variants: YOLO_small, YOLO_tiny, and a YOLO_face model borrowed from another darknet fork.

The interesting bit

The heavy lifting isn’t retraining—it’s weight archaeology. The author extracted values from darknet’s binary .weight files and repacked them into TensorFlow checkpoint format. A converter script (added in 2016) automates that translation, which was genuinely useful when TensorFlow and darknet ecosystems barely talked to each other.

Key highlights

• Supports YOLO_small, YOLO_tiny, and YOLO_face via pretrained weights only
• Includes a darknet-to-TensorFlow weight converter
• Can run as CLI tool or imported as YOLO_TF() class in other scripts
• Outputs to console, image files, or text files with box coordinates and probabilities
• Requires only TensorFlow and OpenCV2

Caveats

• No training support; README explicitly tells you to use darknet for that
• Last meaningful update was February 2017; weights live on Google Drive links of uncertain durability
• License carries a “Do not use this on commercial!” warning from the original darknet terms

Verdict

Worth a look if you’re maintaining legacy TensorFlow 1.x pipelines or studying how early YOLO weights map to TF ops. Everyone else should use modern frameworks—Ultralytics, TensorFlow’s own OD API, or literally anything from this decade.