New 3D Gaussian Splatting Papers Advance Tracking, Expose Adversarial Weaknesses

Overview

• Two arXiv preprints on 3D Gaussian Splatting, posted Thursday, introduce a frequency-based tracking objective and reveal practical adversarial threats to feed-forward models.
• SpectralSplats supervises renders in the frequency domain with global sinusoidal “Spectral Moments” and uses Frequency Annealing to shift from coarse alignment to fine detail.
• By moving away from pixel-overlap losses that produce zero gradients when scenes are misaligned, the approach recovers stable updates even when the render and target do not overlap.
• AdvSplat shows that pretrained, single-pass 3DGS can be derailed by imperceptible pixel changes using white-box attacks and two query-efficient black-box methods that optimize frequency-parameterized noise.
• The authors present these as preprint results that need peer review, and the findings press for robustness research as fast 3D capture spreads to AR/VR, robotics, and commercial pipelines; a separate Apple study the same day reported a new perceptual loss preferred by human raters for crisper 3DGS renders.