The paper "CoherentGS: Sparse Novel View Synthesis with Coherent 3D Gaussians" addresses the challenge of generating high-quality novel views from sparse input images using 3D Gaussian Splatting (3DGS). The authors propose a method to regularize the 3DGS optimization process, ensuring that the Gaussians representing the scene do not overfit to the training images and produce coherent and realistic results. Key contributions include: 1. **Structured Gaussian Representation**: Each pixel in the input images is assigned a single Gaussian, allowing for structured control over the Gaussians in 2D image space. 2. **Coherency Constraints**: Single and multi-view constraints are introduced to enforce coherency in the movement of Gaussians. Single-view constraints use an implicit convolutional decoder to ensure smooth deformation, while multi-view constraints use a total variation loss to maintain smoothness across different views. 3. **Flow-Based Regularization**: A flow-based loss function is introduced to ensure that corresponding pixels in different images have similar Gaussians, enhancing the consistency of the reconstruction. 4. **Depth-Based Initialization**: The position of the Gaussians is initialized using monocular depth estimates, which helps in aligning the representations from different images and improving the optimization process. The method is evaluated on two datasets, LLFF and NVS-RGBD, showing significant improvements over state-of-the-art NeRF-based approaches in terms of quality and speed. The authors demonstrate that their method can produce high-quality texture and geometry, even in challenging scenarios with few input images, and can also handle occluded regions by identifying and inpainting them.The paper "CoherentGS: Sparse Novel View Synthesis with Coherent 3D Gaussians" addresses the challenge of generating high-quality novel views from sparse input images using 3D Gaussian Splatting (3DGS). The authors propose a method to regularize the 3DGS optimization process, ensuring that the Gaussians representing the scene do not overfit to the training images and produce coherent and realistic results. Key contributions include: 1. **Structured Gaussian Representation**: Each pixel in the input images is assigned a single Gaussian, allowing for structured control over the Gaussians in 2D image space. 2. **Coherency Constraints**: Single and multi-view constraints are introduced to enforce coherency in the movement of Gaussians. Single-view constraints use an implicit convolutional decoder to ensure smooth deformation, while multi-view constraints use a total variation loss to maintain smoothness across different views. 3. **Flow-Based Regularization**: A flow-based loss function is introduced to ensure that corresponding pixels in different images have similar Gaussians, enhancing the consistency of the reconstruction. 4. **Depth-Based Initialization**: The position of the Gaussians is initialized using monocular depth estimates, which helps in aligning the representations from different images and improving the optimization process. The method is evaluated on two datasets, LLFF and NVS-RGBD, showing significant improvements over state-of-the-art NeRF-based approaches in terms of quality and speed. The authors demonstrate that their method can produce high-quality texture and geometry, even in challenging scenarios with few input images, and can also handle occluded regions by identifying and inpainting them.