Porting YOLO to TensorFlow, weights and all

What it does Darkflow is a TensorFlow adaptation of the YOLO real-time object-detection network originally built in darknet. It lets you load existing YOLO weights, fine-tune on your own dataset, or train from scratch entirely within the TensorFlow ecosystem. Once you’re happy with the model, it exports a frozen protobuf graph plus a metadata JSON file that can be dropped into mobile apps or C++ pipelines.

The interesting bit The project handles the awkward weight-format translation for you: it can load darknet .weights files, compare layer shapes against a new config, and reuse every identical layer while randomly initializing only the mismatched ones. It also bundles post-processing metadata—anchors, labels, thresholds—into a .meta file alongside the .pb graph so your mobile build doesn’t need to drag the original .cfg around.

Key highlights

• Loads YOLO v1 and v2 weights directly from darknet binaries
• Supports transfer learning by reusing matching layers across different model configs
• Exports frozen TensorFlow protobuf graphs for Java, C++, or Objective-C++ deployment
• Offers a Python API (TFNet) and JSON output with bounding-box coordinates and confidence scores
• Runs real-time detection on images, video files, or webcam streams

Caveats

• Explicitly targets TensorFlow 1.0, which is several major versions behind current releases
• The maintainer is actively looking for help, so expect some issue-tracker dust
• Mobile export instructions reference TensorFlow C++ API docs from the r0.9 era

Verdict A practical stopgap if you need to migrate legacy YOLO v1/v2 weights into TensorFlow or ship frozen graphs to mobile. Otherwise, the documented dependency on TensorFlow 1.0 and YOLO v2 suggests most new projects will want a more current stack.