Recognizing Traffic Lights with Deep Learning

This repo contains the files used to train and run the classifier described in this blog post. This was done during a challenge by Nexar to recognize traffic lights based on images taken by their dashcam app.

Dependencies

Caffe with python bindings.

Directory contents:

/model: contain a caffe deploy.prototxt file and three weights files. The three weights files are used together in a model ensemble.

/testing: has jupyter notebook files that run the model and perform the weighted average.

/training: contains the files needed to train the model (except the training data)

Training the model

The images were first converted to lmdb format and resized to 256x256 using this command:

GLOG_logtostderr=1 ~/caffe/build/tools/convert_imageset \
    --resize_height=256 --resize_width=256 --shuffle  \
    ~/nexar/images/ \
    ~/nexar/labels_test.txt \
    ~/nexar/lmdb/test_lmdb

Each model has a directory in training with some or all of the following files:

solver.prototxt   caffe solver file
solver_p2.prototxt  caffe solver file with lower base learning rate
train_val.prototxt  network training file
rotation_layer.py   python caffe layer for data augmentation with rotation

squeeze_net_manual_scratch__os was training from scratch. The other two models were fine-tuning from weights trained on ImageNet. The weights file is named squeezenet_v1.0.caffemodel.