Flow Compare: Conditional Normalizing Flows for Point Cloud Change Detection

Abstract

Despite significant progress in 3D deep learning for tasks such as classification and semantic segmentation, robust change detection techniques for complex, coloured environments have not been developed. This is in part due to the absence of labelled change detection datasets and the inherent difficulty of constructing such datasets despite the abundance of unlabeled data. Flow Compare is a fully unsupervised approach that leverages expressive generative models with iterative attention trained on multi-temporal coloured point clouds. Change detection is achieved by reframing the problem as anomaly detection given a learnt conditional distribution. Training pairs are formed by co-registered multi-temporal extracts from coloured point cloud scenes. The inherent class imbalance due to the rarity of semantically important change, which is problematic for supervised approaches, is here harnessed to guarantee that relevant changes are considered anomalies under the learnt distribution. This approach shows promise in detecting not only geometric change but also colour change whilst being robust to common semantically unimportant change.