Fine-tuning FasterRCNN to recognize traffic lights

A deep learning model trained on a FasterRCNN backbone and fine-tuned on the LISA dataset. Source code for this project. Working notebook is also hosted on Google Colab.

Items to try in the future

• Experiment with different architectures (YOLO, SSD, etc.)
• Manually collect and annotate some data myself to expand training set
• Optimize inference and training runtimes using various techniques (gradient accumulation, batch normalization, less loops in raw python, etc.)