How robust is multi-view consistency under fast camera motion?

3D Gaussian Splatting and its SLAM descendants assume reasonably smooth camera trajectories. The base 3DGS paper reports its 30+ FPS results on Mip-NeRF 360 captures, where image overlap is high and motion is essentially quasi-static. The capture protocol in the Mip-NeRF 360 dataset is a tripod-paced orbit; nothing in the published results addresses what the representation does when the camera moves fast, accumulates motion blur, or covers a scene with sub-second pans.

The question becomes pointed once you push splatting into the SLAM regime. GS-SLAM and Gaussian Splatting SLAM / MonoGS both report tracking accuracy on the Replica and TUM-RGBD benchmarks — but those benchmarks have well-behaved trajectories. Compare against DROID-SLAM (Teed & Deng, 2021) on EuRoC or TartanAir, where aggressive motion is part of the protocol and deep BA on learned features is the only thing keeping the tracker locked on. The splatting-SLAM systems haven't published numbers on those harder benchmarks, and the few qualitative results suggest the photometric loss struggles when the same Gaussian is seen at radically different scale and exposure across adjacent frames.

The mechanism is plausible. A 3D Gaussian's anisotropic covariance is optimised to match a particular projected footprint in screen space. Fast motion changes that footprint between consecutive frames; motion blur spreads radiance across pixels that weren't part of the primitive's original support. The cumulative effect is that the densification schedule clones primitives to cover the blur, which then have to be pruned when the camera slows down. That dynamic is exactly the kind of thing that the revising-densification line has started to question even in the static case.

What would resolve this: a benchmark that pairs (a) a high-frame-rate capture with known ground-truth poses, (b) a deliberately fast trajectory (sub-second 360° pans, hand-held running motion), and (c) a head-to-head comparison across DROID-SLAM, MonoGS, GS-SLAM, and a NICE-SLAM baseline, on both trajectory accuracy and reconstruction quality. The corpus currently has none of those numbers in any single paper. Until that exists, "splatting SLAM has overtaken classical SLAM" is the kind of claim that only holds inside the easy half of the operating envelope.