Skip to content
Code and files of the deep learning model used to win the Nexar Traffic Light Recognition challenge
Branch: master
Clone or download
davidbrai add script used to relabel model mistakes
This script helped me quickly label images. It goes over images
and opens them on at a time.
Latest commit 14749da Apr 19, 2017
Type Name Latest commit message Commit time
Failed to load latest commit information.
analysis add script used to relabel model mistakes Apr 19, 2017
model initial commit Jan 17, 2017
testing initial commit Jan 17, 2017
training initial commit Jan 17, 2017
LICENSE initial commit Jan 17, 2017 Rename README to Jan 17, 2017

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.


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 \

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   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.

You can’t perform that action at this time.