What it does

YOLO9000 detects objects across 9,000 classes in a single forward pass. This repository bundles the original darknet implementation, pre-trained weights (split into chunks to dodge GitHub’s file limits), and config files so you can run inference on images or video without hunting down scattered dependencies.

The interesting bit

The weights are deliberately chunked with split -b 95m and reassembled with cat — a low-tech workaround for a low-tech problem. The README also walks through GPU compilation, video pipeline setup with ffmpeg, and even points to a Keras converter, making this more of a survival guide than a code dump.

Key highlights

• 9,000 object classes from the CVPR 2017 paper, ready out of the box
• CPU inference works immediately; GPU requires editing a Makefile and ~8GB VRAM
• Includes video demo pipeline: darknet → frame images → ffmpeg → mp4
• Weights reassembly is manual (cat or Windows type), not automated
• Points to YAD2K for Keras weight conversion if you want to escape C

Caveats

• Windows setup requires a hard git reset to a specific darknet commit; latest master needs a submodule pull
• No OpenCV by default, so outputs save to predictions.png instead of displaying live
• “State of the art” claim is from 2017; the model architecture and dependencies are frozen in time

Verdict

Grab this if you need to reproduce YOLO9000 results exactly as published, or you’re teaching historical object detection. Skip it if you want modern accuracy, pip-installable convenience, or anything that runs without editing Makefiles.