Out-of-Distribution Detection in Multi-Label Datasets using Latent Space of \beta-VAE

Learning-enabled components in autonomous systems are widely used for perception tasks but often fail on out-of-distribution inputs not seen during training, creating safety risks. Traditional out-of-distribution detection approaches struggle with multi-label datasets where multiple environmental factors vary simultaneously. This work is innovative because it provides a practical approach using generative models to detect out-of-distribution images in complex scenarios, supporting safe deployment of learning-enabled autonomous systems.

This paper presents methodology for detecting out-of-distribution examples in multi-label datasets using the latent space of beta-VAE models. The approach trains a beta-VAE for each partition of image data with specific generative factor values and uses KL-divergence metrics to identify when test images have factor values not seen during training. The methodology enables detection of safety-critical out-of-distribution scenarios in autonomous systems operating with multiple environmental factors.

The methodology successfully detects out-of-distribution variations in the nuScenes dataset across multiple generative factors including time-of-day, traffic density, and pedestrian presence. Results demonstrate that the approach can identify safety-critical distribution shifts and that appropriately selected beta-VAE models achieve better detection performance than standard approaches. The work shows that generative models can effectively support safety-critical out-of-distribution detection in autonomous systems.