FeedPoinTrS: Integrating Feedback Control in the PoinTr Series for Point Cloud Completion

Point Cloud Completion based on deep learning significantly improves algorithm capability and stability. However, existing networks are predominantly open-loop structures and fail to leverage the bias between the ground truth and the completed point clouds to further enhance network training. In this work, we propose the FeedPoinTrS network, which integrates a feedback system into the PoinTr network series, forming a closed-loop structure. This proposed network enables dynamic adjustment of model parameters, allowing it to better adapt to varying input data characteristics. Extensive experiments on the standard PCN, ShapeNet-55, and ShapeNet-34 datasets demonstrate the superiority of the FeedPoinTrS network in point cloud completion. On the ShapeNet-34 dataset, our FeedAdaPoinTr surpasses AdaPoinTr by 4.11% on the seen test set and 8.13% on the unseen test set under Chamfer Distance-L2(CD-L2). Additionally, FeedPoinTr outperforms PoinTr by 38.21% and 32.68%, respectively, on the same metrics.