The paper introduces Convolutional Occupancy Networks (COnNs), a novel approach for 3D reconstruction that combines the strengths of convolutional neural networks (CNNs) and implicit representations. COnNs address the limitations of existing implicit methods, which often struggle with complex or large-scale scenes due to their fully-connected network architecture. By integrating convolutional encoders with implicit occupancy decoders, COnNs incorporate inductive biases and enable structured reasoning in 3D space. The key contributions of the paper include: 1. **Model Architecture**: COnNs use convolutional operations to achieve translation equivariance, allowing for the integration of local and global information. The model processes 3D inputs (e.g., noisy point clouds or coarse voxel grids) through a task-specific neural network, followed by convolutional encoding and decoding to predict occupancy probabilities. 2. **Evaluation**: The effectiveness of COnNs is demonstrated through experiments on complex geometry reconstruction from noisy point clouds and low-resolution voxel representations. The method shows superior performance in terms of accuracy, generalization to real data, and handling of large indoor scenes. 3. **Contributions**: The paper identifies the limitations of current implicit 3D reconstruction methods and proposes a flexible translation equivariant architecture that enables accurate 3D reconstruction from single objects to large scenes. It also shows that the model generalizes well from synthetic to real scenes and novel object categories. 4. **Related Work**: The paper reviews existing methods for 3D reconstruction, including volumetric representations, point clouds, mesh-based representations, and implicit representations. It highlights the key limitations of these approaches and how COnNs address them. 5. **Methodology**: The paper details the encoder, decoder, occupancy prediction, and training procedures of the COnNs model. It explores different feature aggregation strategies and interpolation techniques to optimize performance and memory efficiency. 6. **Experiments**: The paper evaluates COnNs on various datasets, including ShapeNet for object-level reconstruction, a synthetic indoor scene dataset for scene-level reconstruction, and real-world datasets like ScanNet v2 and Matterport3D. Results show that COnNs outperform state-of-the-art methods in terms of accuracy, robustness to noise, and ability to handle large scenes. 7. **Conclusion**: The paper concludes by discussing the advantages of COnNs, including their ability to generalize to unseen classes, novel room layouts, and large-scale indoor spaces. It also highlights future directions for applying the novel representation to other domains such as implicit appearance modeling and 4D reconstruction.The paper introduces Convolutional Occupancy Networks (COnNs), a novel approach for 3D reconstruction that combines the strengths of convolutional neural networks (CNNs) and implicit representations. COnNs address the limitations of existing implicit methods, which often struggle with complex or large-scale scenes due to their fully-connected network architecture. By integrating convolutional encoders with implicit occupancy decoders, COnNs incorporate inductive biases and enable structured reasoning in 3D space. The key contributions of the paper include: 1. **Model Architecture**: COnNs use convolutional operations to achieve translation equivariance, allowing for the integration of local and global information. The model processes 3D inputs (e.g., noisy point clouds or coarse voxel grids) through a task-specific neural network, followed by convolutional encoding and decoding to predict occupancy probabilities. 2. **Evaluation**: The effectiveness of COnNs is demonstrated through experiments on complex geometry reconstruction from noisy point clouds and low-resolution voxel representations. The method shows superior performance in terms of accuracy, generalization to real data, and handling of large indoor scenes. 3. **Contributions**: The paper identifies the limitations of current implicit 3D reconstruction methods and proposes a flexible translation equivariant architecture that enables accurate 3D reconstruction from single objects to large scenes. It also shows that the model generalizes well from synthetic to real scenes and novel object categories. 4. **Related Work**: The paper reviews existing methods for 3D reconstruction, including volumetric representations, point clouds, mesh-based representations, and implicit representations. It highlights the key limitations of these approaches and how COnNs address them. 5. **Methodology**: The paper details the encoder, decoder, occupancy prediction, and training procedures of the COnNs model. It explores different feature aggregation strategies and interpolation techniques to optimize performance and memory efficiency. 6. **Experiments**: The paper evaluates COnNs on various datasets, including ShapeNet for object-level reconstruction, a synthetic indoor scene dataset for scene-level reconstruction, and real-world datasets like ScanNet v2 and Matterport3D. Results show that COnNs outperform state-of-the-art methods in terms of accuracy, robustness to noise, and ability to handle large scenes. 7. **Conclusion**: The paper concludes by discussing the advantages of COnNs, including their ability to generalize to unseen classes, novel room layouts, and large-scale indoor spaces. It also highlights future directions for applying the novel representation to other domains such as implicit appearance modeling and 4D reconstruction.