The statement is discussing the evaluation of an Occupancy Network (ONet) in comparison to a voxel representation at various resolutions. The volumetric IoU (Intersection over Union) is measured to evaluate how accurately the ONet represents the ground truth mesh. The results show that the ONet can represent the entire dataset with high accuracy, while a low-resolution voxel representation cannot. Additionally, the ONet can encode all training samples with as little as 6 million parameters, while voxel representations require increasingly more memory with resolution. The qualitative results also show that the ONet can represent details of the 3D geometry that are lost in a low-resolution voxelization.
