The article presents a new method called Point Regress AutoEncoder (Point-RAE) for regressive autoencoders in point cloud self-supervised learning. The method decouples functions between the decoder and encoder by introducing a mask regressor that predicts the masked patch representation from the visible patch representation. The decoder then reconstructs the target from the predicted masked patch representation. The method also introduces an alignment constraint to ensure representations for masked patches are aligned with the masked patch presentations computed from the encoder. The method has demonstrated high accuracy and efficiency in pre-training and generalizes well on various downstream tasks.
Publication date: 25 Sep 2023
Project Page: https://github.com/liuyyy111/Point-RAE
Paper: https://arxiv.org/pdf/2310.03670
