Zero-shot anomalous object detection using unsupervised metric learning

This paper proposes a novel unsupervised metric learning approach to detect anomalous/novel objects. Existing object detection approaches either cannot detect novel categories or require human annotations even in a small scale (such as few-shot learning). To overcome this, especially for robotic applications where human annotations are not available, this project leverages unsupervised representation learning and unsupervised metric learning to discover feature prototypes of unknown fine-grained categories i.e. clusters in the low-dimensional embedding space. Specifically, the proposed approach leverages deep clustering and self-reconstruction to learn feature prototypes for normal objects. More importantly, we interpolate the latent features and generate pseudo anomalous examples to learn the embedding space of good compactness and sparseness to learn a discriminative embedding space, facilitating distinguishing anomalous examples from normal ones. The learned prototypes can be further used to infer the probability of novel objects using the metric distance to the prototypes. The proposed unsupervised learning approach is also integrated with a Region Proposal Network as a detection pipeline and real-time detection is achieved. This paper uses the StreetHazards dataset of CAOS benchmark for training and evaluation and comparison experiments are implemented to demonstrate the effectiveness of the proposed approach.