Training Yolo on LEGO

This repository contains experiments of transfer learning using YOLO on a new synthetical LEGO data set ROUGH AND UNDOCUMENTED!

Prerequisite

• Tensorflow

Organisation

• dumps_weights_to_pkl.py save tiny yolo weights obtained from darkflow to a pkl
• train.py use those weights to retrain the network defined in tiny_yolo.py
• load.py contains function to save and create Tensorflow graphs including a freeze_graph(folder) that makes a graph as a tensorflow checkpoint reusable
• test.py apply a 'frozen graph' to a test image

References and Acknowledgements

This code is based on the foillowing paper Redmon, Joseph, et al. "You only look once: Unified, real-time object detection." arXiv preprint arXiv:1506.02640 (2015). website : http://pjreddie.com/darknet/yolo/

Some python functions and are taken from or modified from : https://github.com/thtrieu/darkflow