The paper introduces 3D Recurrent Reconstruction Neural Network (3D-R2N2), a novel recurrent neural network architecture for 3D object reconstruction from single and multi-view images. Inspired by shape priors, 3D-R2N2 learns a mapping from images to 3D shapes using a large collection of synthetic data. The network can handle objects observed from arbitrary viewpoints and outputs a 3D occupancy grid without requiring image annotations or object class labels. Key contributions include an extension of LSTM to accommodate multi-view inputs, unifying single- and multi-view reconstruction, minimal supervision during training and testing, and superior performance in single-view reconstruction and challenging scenarios where traditional methods fail due to lack of texture or wide baselines. The network's ability to selectively update hidden representations and handle self-occlusions is highlighted, along with its effectiveness in various datasets, including PASCAL 3D, ShapeNet, and Online Products.The paper introduces 3D Recurrent Reconstruction Neural Network (3D-R2N2), a novel recurrent neural network architecture for 3D object reconstruction from single and multi-view images. Inspired by shape priors, 3D-R2N2 learns a mapping from images to 3D shapes using a large collection of synthetic data. The network can handle objects observed from arbitrary viewpoints and outputs a 3D occupancy grid without requiring image annotations or object class labels. Key contributions include an extension of LSTM to accommodate multi-view inputs, unifying single- and multi-view reconstruction, minimal supervision during training and testing, and superior performance in single-view reconstruction and challenging scenarios where traditional methods fail due to lack of texture or wide baselines. The network's ability to selectively update hidden representations and handle self-occlusions is highlighted, along with its effectiveness in various datasets, including PASCAL 3D, ShapeNet, and Online Products.