Abstract
In this paper, we present a Scale-adaptive method for Anti-aliasing GaussianSplatting (SA-GS). While the state-of-the-art method Mip-Splatting needsmodifying the training procedure of Gaussian splatting, our method functions attest-time and is training-free. Specifically, SA-GS can be applied to anypretrained Gaussian splatting field as a plugin to significantly improve thefield's anti-alising performance. The core technique is to apply 2Dscale-adaptive filters to each Gaussian during test time. As pointed out byMip-Splatting, observing Gaussians at different frequencies leads to mismatchesbetween the Gaussian scales during training and testing. Mip-Splatting resolvesthis issue using 3D smoothing and 2D Mip filters, which are unfortunately notaware of testing frequency. In this work, we show that a 2D scale-adaptivefilter that is informed of testing frequency can effectively match the Gaussianscale, thus making the Gaussian primitive distribution remain consistent acrossdifferent testing frequencies. When scale inconsistency is eliminated, samplingrates smaller than the scene frequency result in conventional jaggedness, andwe propose to integrate the projected 2D Gaussian within each pixel duringtesting. This integration is actually a limiting case of super-sampling, whichsignificantly improves anti-aliasing performance over vanilla GaussianSplatting. Through extensive experiments using various settings and bothbounded and unbounded scenes, we show SA-GS performs comparably with or betterthan Mip-Splatting. Note that super-sampling and integration are only effectivewhen our scale-adaptive filtering is activated. Our codes, data and models areavailable at https://github.com/zsy1987/SA-GS.