New 3D Gaussian Splatting Papers Advance Tracking and Expose Attack Risks

Overview

• Two arXiv preprints released Thursday introduce a frequency-based fix for failed tracking and a study that reveals adversarial weak points in feed-forward 3DGS.
• 3D Gaussian Splatting renders scenes from many tiny 3D blobs in real time, which makes fast, sharp views possible but can stall tracking when the render and the target do not overlap.
• SpectralSplats trains in the frequency domain using global sinusoid features called Spectral Moments and a Frequency Annealing schedule that moves from rough alignment to fine detail even with large misalignment.
• AdvSplat finds that feed-forward 3DGS can be derailed by imperceptible input changes, using white-box attacks and two query-efficient black-box methods that search pixel tweaks through a frequency-based parameterization.
• The results are preprints and need validation, and they arrive as teams push 3DGS forward, with Apple researchers the same day reporting a perceptual loss (WD-R) that human raters preferred for sharper textures without more compute.