The paper introduces 3D ShapeNets, a Convolutional Deep Belief Network (CDBN) designed to represent 3D shapes as probability distributions of binary variables on a 3D voxel grid. This approach aims to enhance object recognition and view planning in 2.5D depth maps, leveraging the recent advancements in inexpensive 2.5D depth sensors like Microsoft Kinect. The model is trained using a large-scale 3D computer graphics dataset, ModelNet, which includes 127,915 3D CAD models. The paper details the architecture of the CDBN, its training process, and its performance in object recognition and next-best-view prediction tasks. Experiments demonstrate that 3D ShapeNets outperforms state-of-the-art methods in 2.5D depth-based object recognition and effectively predicts the best next view to reduce recognition uncertainty. The paper also discusses the construction of ModelNet and the evaluation of 3D ShapeNets on various datasets, showing its robustness and generalization capabilities.The paper introduces 3D ShapeNets, a Convolutional Deep Belief Network (CDBN) designed to represent 3D shapes as probability distributions of binary variables on a 3D voxel grid. This approach aims to enhance object recognition and view planning in 2.5D depth maps, leveraging the recent advancements in inexpensive 2.5D depth sensors like Microsoft Kinect. The model is trained using a large-scale 3D computer graphics dataset, ModelNet, which includes 127,915 3D CAD models. The paper details the architecture of the CDBN, its training process, and its performance in object recognition and next-best-view prediction tasks. Experiments demonstrate that 3D ShapeNets outperforms state-of-the-art methods in 2.5D depth-based object recognition and effectively predicts the best next view to reduce recognition uncertainty. The paper also discusses the construction of ModelNet and the evaluation of 3D ShapeNets on various datasets, showing its robustness and generalization capabilities.