ObsGS: Observability-Gated Densification for Large-Scene Gaussian Reconstruction

Large-scene Gaussian reconstruction must spend a limited primitive budget on centres the cameras can triangulate. Standard adaptive density control instead clones or splits where the view-space photometric gradient is large. On a forward city road the two decisions diverge: many cameras share one bearing, the image residual can still be large, and the depth of a new centre stays weakly determined. We introduce ObsGS, which multiplies that photometric score by the scale-normalized smallest eigenvalue of the multiview projection information matrix. The renderer, photometric loss, and optimizer are unchanged, and training does not read depth. On two MatrixCity street acquisitions, with ground-truth cameras, a shared 2,000-step schedule, and a matched cap of 10,000 Gaussians, ObsGS raises centre-to-surface F-score at $2$ m from 0.099 to 0.143 on the road and from 0.245 to 0.293 on the vertical sequence, above a view-count gate in both cases. The advantage is a camera-only control that moves a fixed budget toward recoverable surfaces. A bearing-angle gate captures much of the same surface movement, and on the vertical sequence ObsGS lowers held-out PSNR by $2.32$ dB, so the measured gain is directional observability with a bounded image-quality cost.