Recognizing Traffic Lights With Deep Learning
summary
Built a traffic-light classifier that wins the challenge with 94.955% (rounded to 95%) accuracy and a 7.84 MB model, trained in Caffe on Amazon GPU instances. Starts from SqueezeNet fine-tuned (also compared to GoogleNet), adds rotational averaging and 5-crop averaging at inference, uses low-learning-rate fine-tuning and ensembling of three models, and manually relabels ~300 wrongly labeled images to boost validation from ~92% to 94.1% — practical, size-constrained, and working.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:50Competition Task
- 4:03Score
- 6:06What I use
- 8:36Results
- 9:32Baseline
- 10:58SqueezeNet
- 16:56Cafe
- 18:05Crop Sampling
- 21:21Overfitting
- 25:22Models
related talks
33mRunning Machine Learning in Production - a Journey to Success
26mSelf Driving cars with Opencv
28mBachelor theses in Deep Learning: Submitted to an Armenian University
41mPractical Aspects of Applying Modern Computer Vision Technologies
59mDeep Neural Networks for Video Applications
18mRecognition and Localization of Parking Signs using Deep Learning
40mHow to Implement a Production Ready Deep Learning Model in 12 Weeks
57mTensorflow, deep learning and modern convolutional neural nets, without a PhD