davidbrai/deep-learning-traffic-lights

Implements a three-model ensemble approach using Caffe with SqueezeNet architecture, combining one model trained from scratch with two ImageNet fine-tuned variants for improved accuracy. Features custom Python rotation layers for data augmentation and processes training data through LMDB format conversion with 256x256 resizing. Predictions are generated via weighted averaging of the ensemble members, with full training pipelines and Jupyter notebooks provided for reproducibility.

483 stars. No commits in the last 6 months.