Wang Hanchen ; Liu Qi ; Yue Xiangyu ; Lasenby Joan ; Kusner Matthew J.

Abstract

We describe a simple pre-training approach for point clouds. It works inthree steps: 1. Mask all points occluded in a camera view; 2. Learn anencoder-decoder model to reconstruct the occluded points; 3. Use the encoderweights as initialisation for downstream point cloud tasks. We find that evenwhen we construct a single pre-training dataset (from ModelNet40), thispre-training method improves accuracy across different datasets and encoders,on a wide range of downstream tasks. Specifically, we show that our methodoutperforms previous pre-training methods in object classification, and bothpart-based and semantic segmentation tasks. We study the pre-trained featuresand find that they lead to wide downstream minima, have high transformationinvariance, and have activations that are highly correlated with part labels.Code and data are available at: https://github.com/hansen7/OcCo