Toward Unsupervised 3d Point Cloud Anomaly Detection Using Variational Autoencoder

Abstract

In this paper, we present an end-to-end unsupervised anomaly detection framework for 3D point clouds. To the best of our knowledge, this is the first work to tackle the anomaly detection task on a general object represented by a 3D point cloud. We propose a deep variational autoencoder based unsupervised anomaly detection network adapted to the 3D point cloud and an anomaly score specifically for 3D point clouds.To verify the effectiveness of the model, we conducted extensive experiments on ShapeNet dataset. Through quantitative and qualitative evaluation, we demonstrate that the proposed method outperforms the baseline method.

3DOS: Towards 3D Open Set Learning -- Benchmarking and Understanding Semantic Novelty Detection on Point Clouds

“…We select two methods from this group. We test a VAE model with reconstruction-based scoring by following one of the few existing works on 3D OOD detection [32]. This is the only unsupervised approach in our analysis: it separates known from unknown samples, but cannot provide category predictions on the known classes samples.…”

“…Up to our knowledge 3D OOD detection and Open Set problems have been studied only by a handful of works. A VAE approach for reconstruction-based 3D OOD detection is provided in [32], together with an analysis on seven classes of the ShapeNet dataset [6], each used in turn as unknown. The study considers different VAE normality scores but does not compare with other baselines.…”

“…It is composed of graph-convolutional and MLP layers, while the decoder is inspired by FoldingNet [57]. Still, VAE was used in one of the few existing 3D OOD detection works [32], so we decided to keep it in our analysis as a reference.…”

“…Density and Reconstruction Based Methods. For our VAE experiments we use the original code publicly released by the authors 2 [32], as well as their same choice on point cloud cardinality (2048 points) and hyperparameters. The encoder is composed of graph-convolutional layers: it takes as input a point cloud and outputs two 512-dimensional vectors representing the mean and variance.…”

“…infancy, with only a small number of works which have just started to scratch the surface of the problem by focusing on particular sub-settings [32,3]. With this work, we draw the community's attention to Open Set 3D learning, which entails developing models designed to process 3D point clouds that can recognize test samples from a set of known categories while avoiding prediction for samples from unknown classes.…”