VariationalVision

An anomaly detection system for self driving cars. I collected around 50,000 images from CARLA autopilot and augmented them to 100,000, then trained a convolutional variational autoencoder to spot unusual driving scenes. It is meant to surface the edge cases that rarely show up in normal driving data, which is useful for places with road conditions that standard datasets miss.