“On Scaling Up 3D Gaussian Splatting Training” explores the potential of training high-parameter 3D Gaussian Splatting (3DGS) models on large-scale, high-resolution datasets. This research addresses the challenges associated with scaling up 3DGS models to handle more complex scenes with higher spatial resolution and larger datasets, which are essential for achieving high-quality 3D scene reconstruction.
The authors introduce RetinaGS, a general model parallel training method for 3DGS. This method employs a proper rendering equation and can be applied to any scene with an arbitrary distribution of Gaussian primitives. RetinaGS enables the exploration of the scaling behavior of 3DGS in terms of the number of primitives and training resolutions, which were previously difficult to investigate. The method demonstrates a clear positive trend in visual quality with an increasing number of primitives.
One of the key innovations of this work is the ability to train a 3DGS model with more than one billion primitives on the full MatrixCity dataset, achieving promising visual quality. The paper highlights the significant computational and memory costs associated with high-resolution training, which necessitates a distributed training approach to make the process feasible.
To achieve efficient distributed training, the authors propose dividing the model space into a set of convex subspaces and assigning subsets of splats to each subspace. This approach allows for the transformation of the original iterative alpha-blending process into a hierarchical form. By computing partial color and alpha values for each subset and then merging these values, the method retains equivalence to the existing single GPU training scheme while enabling effective scaling.
The paper provides extensive experimental results to demonstrate the effectiveness of RetinaGS. The authors evaluate their approach on several datasets and compare it with existing state-of-the-art techniques. The results show that RetinaGS consistently improves the visual quality of rendered views, even at high resolutions and with large data volumes. In conclusion, “On Scaling Up 3D Gaussian Splatting Training” presents a significant advancement in the field of 3D scene reconstruction. By introducing RetinaGS, the authors offer a scalable and efficient method for training high-parameter 3DGS models on large-scale datasets. This research has important implications for various applications, including virtual reality, gaming, and digital content creation, where high-quality 3D scene reconstruction is crucial.
