VariationalVision

Tech Stack:

PyTorch, CARLA simulator, ResNet, A100 GPU (training), RTX 3050 Ti (inference)

Goal:

Build an anomaly detection system that identifies unusual driving scenarios to improve self-driving car training datasets.

What I did:

• •Collected 50,000 images using CARLA autopilot mode, augmented to 100,000 samples
• •Built a convolutional variational autoencoder with KL divergence and reconstruction loss
• •Used ResNet for classification of anomalous scenes
• •Applied kernel density estimation and spatial downsampling for feature reduction

What was achieved:

System successfully distinguishes normal from anomalous driving scenes. Useful for capturing edge cases in regions like South Asia with different road conditions.