YOLOv3-on-LISA-Traffic-Sign-Detection-with-darknet

This project is to improve YOLOv3 performance by GIOU instead of IOU and the integration of conv and batch_normalization layers

Dependence

1. Ubuntu 16.04
2. darknet
3. Python 3.5

Dataset

LISA Traffic Sign Dataset: Laboratory for Intelligent and Safe Automobiles Link: http://cvrr.ucsd.edu/LISA/lisa-traffic-sign-dataset.html

Implementation

1. Download darknet framework Link: https://github.com/pjreddie/darknet
2. Download pretrained model weight Link: https://pjreddie.com/media/files/darknet53.conv.74
3. Convert the data format of LISA dataset by pip install Jinja2-2.10.1-py2.py3-none-any.whl and parse_lisa.py on LISA_TS/allAnnotations.csv
4. Add data/voc-lisa.names
5. Add cfg/voc-lisa.data
6. Add cfg/yolov3-voc-lisa-giou.cfg with arg set up
7. For the integration of conv and batch_normalization layers, define fuse_conv_batchnorm(network net) in include/darknet.h and src/network.c, and call it in src/demo.c and examples/detector.c
8. For given the choice between optimizing a metric itself vs. a surrogate loss function, the optimal choice is the metric itself. The usage of GIOU insteads of l-1 and l-2 norms
9. ./darknet detector train cfg/voc-lisa.data cfg/yolov3-voc-lisa-giou.cfg darknet53.conv.74 2>1

TODO

1. Train model with LISA by CuDNN and OpenCV
2. The visualization of training performance
3. The usage of Clustering algorithms on Anchor Box size
4. Compress YOLOv3
5. Convert trained model to Caffe model for the deployment of Embedded system

Reference

1. Darknet: Open Source Neural Networks in C, https://pjreddie.com/darknet/
2. Generalized Intersection over Union: a Metric and a Loss for Bounding Box Regression, https://arxiv.org/pdf/1902.09630.pdf