Point cloud completion—filling in missing geometry from partial 3D scans—is usually tackled by deep networks that pile on attention modules. The authors of a new arXiv preprint argue that this design is both overcomplicated and incomplete: it relies entirely on feed-forward connections, with no feedback from deeper layers. They take inspiration from the brain, where recurrent feedforward-feedback loops in the visual cortex generate illusory contours, a mechanism the brain uses to infer missing visual information.
Building on predictive coding, the team introduces BRICE, a plug-and-play encoder that adds such recurrent loops to existing completion networks. The module is meant to 'imagine' the missing parts of a point cloud rather than simply transform the visible input. Because it is modular, it can be inserted into mainstream baselines without redesigning them.
Experiments on the MVP2K, PCN, and PlantPCom datasets show that BRICE improves completion performance across several baselines. The paper includes code, and the authors frame the work as a step toward more interpretable, unified network designs. As a single-source report, these results have not yet been independently replicated.