This paper introduces 3D ShapeNets, a convolutional deep belief network that represents 3D shapes as a probability distribution of binary variables on a 3D voxel grid. The model is designed for 2.5D object recognition and next-best-view prediction. It uses a large-scale 3D computer graphics dataset called ModelNet, containing 127,915 3D CAD models, to train the model. The model can recognize objects from 2.5D depth maps and predict the best next view to reduce recognition uncertainty. The model is trained using a layer-wise pre-training approach followed by generative fine-tuning. It is evaluated on the NYU dataset and shows superior performance in 3D shape classification and retrieval compared to other state-of-the-art methods. The model also supports next-best-view planning for object recognition, which is crucial for reducing uncertainty when recognition is ambiguous. The paper also presents experiments on view-based 2.5D recognition and next-best-view prediction, demonstrating the effectiveness of the model in real-world scenarios. The results show that the model can accurately recognize objects and predict the best next view for recognition. The model is trained on a large-scale 3D CAD dataset and is capable of handling a wide range of object categories. The paper concludes that 3D ShapeNets is a powerful tool for 3D shape representation and object recognition, and future work includes constructing a large-scale Kinect-based 2.5D dataset for further evaluation.This paper introduces 3D ShapeNets, a convolutional deep belief network that represents 3D shapes as a probability distribution of binary variables on a 3D voxel grid. The model is designed for 2.5D object recognition and next-best-view prediction. It uses a large-scale 3D computer graphics dataset called ModelNet, containing 127,915 3D CAD models, to train the model. The model can recognize objects from 2.5D depth maps and predict the best next view to reduce recognition uncertainty. The model is trained using a layer-wise pre-training approach followed by generative fine-tuning. It is evaluated on the NYU dataset and shows superior performance in 3D shape classification and retrieval compared to other state-of-the-art methods. The model also supports next-best-view planning for object recognition, which is crucial for reducing uncertainty when recognition is ambiguous. The paper also presents experiments on view-based 2.5D recognition and next-best-view prediction, demonstrating the effectiveness of the model in real-world scenarios. The results show that the model can accurately recognize objects and predict the best next view for recognition. The model is trained on a large-scale 3D CAD dataset and is capable of handling a wide range of object categories. The paper concludes that 3D ShapeNets is a powerful tool for 3D shape representation and object recognition, and future work includes constructing a large-scale Kinect-based 2.5D dataset for further evaluation.