GridFormer: A Point-Grid Transformer for Surface Reconstruction GridFormer is a novel and efficient attention mechanism that bridges the gap between space and point clouds for surface reconstruction. It introduces a Point-Grid Transformer (GridFormer) to leverage both point and grid features, enhancing spatial expressiveness while maintaining computational efficiency. The method uses a two-stage training strategy incorporating margin binary cross-entropy loss and boundary sampling to improve reconstruction accuracy. GridFormer outperforms state-of-the-art methods on widely used benchmarks, producing more precise geometry reconstructions. The method addresses the challenges of reconstructing continuous surfaces from discrete point clouds by using a point-grid attention mechanism. It treats the grid as a transfer point connecting space and point clouds, enabling the network to learn the relationship between input and grid features. This approach allows for more accurate representation of object structures and better handling of boundary issues. GridFormer is evaluated on the ShapeNet and Synthetic Rooms datasets, demonstrating its effectiveness in both object-level and scene-level reconstruction. It achieves more detailed and accurate reconstructions compared to other methods, particularly in capturing fine structures. The method also shows strong generalization performance on the ScanNet-v2 dataset, reconstructing smoother and more complete surfaces. The method's contributions include the introduction of GridFormer for surface reconstruction, a two-stage training strategy with margin binary cross-entropy loss and boundary sampling, and validation through both object-level and scene-level experiments. GridFormer is efficient, with lower computational time and GPU memory usage compared to other methods, and it achieves comparable or better results. The method is implemented in PyTorch and uses the Adam optimizer. It is trained on the Synthetic Rooms dataset and tested on ScanNet-v2, with results showing improved performance in terms of volumetric IoU, Chamfer distance, normal consistency, and F-Score. The method also includes an ablation study on grid representation, downsampling, and boundary optimization, demonstrating its effectiveness in reducing error bounds and improving reconstruction accuracy.GridFormer: A Point-Grid Transformer for Surface Reconstruction GridFormer is a novel and efficient attention mechanism that bridges the gap between space and point clouds for surface reconstruction. It introduces a Point-Grid Transformer (GridFormer) to leverage both point and grid features, enhancing spatial expressiveness while maintaining computational efficiency. The method uses a two-stage training strategy incorporating margin binary cross-entropy loss and boundary sampling to improve reconstruction accuracy. GridFormer outperforms state-of-the-art methods on widely used benchmarks, producing more precise geometry reconstructions. The method addresses the challenges of reconstructing continuous surfaces from discrete point clouds by using a point-grid attention mechanism. It treats the grid as a transfer point connecting space and point clouds, enabling the network to learn the relationship between input and grid features. This approach allows for more accurate representation of object structures and better handling of boundary issues. GridFormer is evaluated on the ShapeNet and Synthetic Rooms datasets, demonstrating its effectiveness in both object-level and scene-level reconstruction. It achieves more detailed and accurate reconstructions compared to other methods, particularly in capturing fine structures. The method also shows strong generalization performance on the ScanNet-v2 dataset, reconstructing smoother and more complete surfaces. The method's contributions include the introduction of GridFormer for surface reconstruction, a two-stage training strategy with margin binary cross-entropy loss and boundary sampling, and validation through both object-level and scene-level experiments. GridFormer is efficient, with lower computational time and GPU memory usage compared to other methods, and it achieves comparable or better results. The method is implemented in PyTorch and uses the Adam optimizer. It is trained on the Synthetic Rooms dataset and tested on ScanNet-v2, with results showing improved performance in terms of volumetric IoU, Chamfer distance, normal consistency, and F-Score. The method also includes an ablation study on grid representation, downsampling, and boundary optimization, demonstrating its effectiveness in reducing error bounds and improving reconstruction accuracy.