ImageNet classification with deep convolutional neural networks
Alex Krizhevsky, Ilya Sutskever et al.
0
Citations
0
Influential Citations
—
Venue
2024
Year
3D Gaussian Splatting (3DGS) has recently revolutionized the field of neural rendering with its high fidelity and efficiency. However, 3DGS heavily depends on the initialized point cloud …
3D Gaussian Splatting (3DGS) has emerged as a powerful alternative to NeRF for real-time neural rendering, offering high fidelity and fast training. However, its performance heavily depends on the quality of the initial point cloud, typically generated by Structure-from-Motion (SfM). In many real-world scenarios, SfM point clouds are sparse or incomplete, leading to artifacts and degraded rendering quality. GaussianPro addresses this critical limitation by introducing a progressive propagation mechanism that actively grows Gaussian primitives into under-covered regions, reducing the reliance on dense initialization.
This paper is significant because it tackles a practical bottleneck in 3DGS adoption. By making the method more robust to imperfect inputs, it paves the way for broader use in dynamic environments, large-scale scenes, and casual capture settings where SfM may fail. The progressive propagation approach also aligns with a trend toward self-supervised and adaptive scene representation learning.
While the abstract does not provide specific numerical metrics, the paper claims improved rendering fidelity and visual quality over standard 3DGS, particularly in scenarios with sparse or incomplete initial point clouds. The method is evaluated on standard benchmarks, likely including Mip-NeRF360, Tanks & Temples, and DTU, though exact PSNR/SSIM/LPIPS values are not available in the abstract. The qualitative improvements are most pronounced in regions where the initial point cloud has gaps.
GaussianPro contributes to the ongoing evolution of 3DGS by making it more robust and practical. Its progressive propagation technique could inspire further research into adaptive scene representations that do not rely on dense priors. This is particularly relevant for real-time applications such as augmented reality, virtual reality, and autonomous navigation, where capturing high-quality point clouds is often challenging. The work also opens avenues for combining 3DGS with learning-based depth estimation or multi-view stereo to further reduce initialization requirements.
Alex Krizhevsky, Ilya Sutskever et al.
Ashish Vaswani, Noam Shazeer et al.
Douglas M. Bates, Martin Mächler et al.
Diederik P. Kingma, Jimmy Ba