SCOUT-SLAM: Structurally-Coupled Dual Uncertainty-Aware 3DGS SLAM in the Wild
Kumaran Karthik, Pramat Shastri Jois, Suresh Sundaram
cs.CV
2026-09-14
SCOUT-SLAM reads tracking uncertainty from multi-view DINOv2 consistency, reaching 1.42 cm mean TUM ATE versus 1.51 cm for WildGS-SLAM.
Uncertainty-aware 3DGS-SLAM (track the camera while a set of 3D Gaussians serves as a renderable map) tries to keep moving people out of the map without a closed set of object classes. WildGS-SLAM learns that uncertainty from rendering residuals: if the current Gaussians cannot explain a pixel, the pixel is treated as dynamic, and tracking then reweights pose residuals with the same mask.
That coupling becomes a loop. Fast camera motion and cluttered foregrounds make the reconstruction unstable, residuals blow up on static walls, tracking trusts a bad mask, poses get worse, and the map gets worse still. Semantic masks can dodge the loop, but they fail on open-set motion.
SCOUT-SLAM, from the Indian Institute of Science, estimates tracking and mapping uncertainty from one shared base network, with different training signals.
The base, MUNet, is a two-layer MLP (hidden width 64) that reads frozen DINOv2 features, a per-pixel field from a self-supervised vision transformer, and predicts mapping uncertainty. Tracking does not train a second MLP. A Tracking Uncertainty Adapter injects LoRA (a pair of low-rank matrices, rank 8, about 5.1k parameters) into MUNet and updates only those matrices at each bundle-adjustment step. The loss is multi-view feature consistency: warp DINOv2 features between observed frames with the current rigid geometry, and treat cosine disagreement as dynamics. That quantity never goes through the current Gaussian render, so tracking uncertainty is not a slave of map quality.
The adapter output becomes a confidence \(w=\mathrm{clamp}(1/u{\mathrm{track}}^2,\epsilon,1)\) that inflates optical-flow covariance on dynamic pixels inside dense BA. Mapping uses a spatially adaptive prior: high tracking confidence forbids inflating uncertainty on static structure; a residual gate relaxes the penalty when the leftover error looks like sensor noise, so Gaussians are not forced to fit flicker. Render loss is divided by detached mapping uncertainty, so only static regions update the map. Metric3D v2 supplies a monocular depth prior. After a 15-keyframe warmup and 1050 MUNet steps, the initial graph is re-optimized with TUA masks.
The rank bottleneck is deliberate. Motion inside a local BA window is low-dimensional; a full MLP overfits transient alignments. Starting from the mapping features also lets the tracker lock onto a newly entering person in fewer iterations than DROID-W, which learns an affine layer from scratch.
ATE RMSE, lower is better. Input-view PSNR is computed after SAM3 masks dynamic objects.
| Method | TUM (cm) | Bonn (cm) | MoCap (cm) | TUM PSNR |
| SCOUT-SLAM | 1.42 | 2.29 | 0.47 | 19.34 |
| WildGS-SLAM | 1.51 | 2.47 | 0.46 | 18.91 |
| DROID-W | 1.52 | 2.32 | 0.65 | n/a |
The largest TUM gain is on walking sequences (fr3/wr 2.91 cm vs WildGS-SLAM 3.15 cm). Bonn gaps are widest on balloon and crowd2. On Wild-SLAM MoCap, SCOUT averages 0.47 cm, a hair behind WildGS-SLAM at 0.46 cm. A standalone MLP with a similar parameter count scores 5.36 cm ATE on Bonn; TUA at rank 8 scores 2.29 cm. Dropping the adaptive prior moves TUM ATE from 1.42 to 1.45 cm and PSNR from 19.34 to 19.20 dB.
For monocular 3DGS-SLAM in rooms with people, this is a structural patch on WildGS-SLAM: tracking uncertainty comes from feature consistency across observations, mapping uncertainty still comes from residuals, and tracking confidence pins down static regions. The gains are millimetres to sub-centimetres, consistent across three dynamic benchmarks. Code is public. It does not replace a segmenter, and it does not solve slow objects hugging the image border.
There is no dedicated Limitations section. On MoCap tabletracking, a slow foreground near the border makes the MFC signal ambiguous, and SCOUT trails WildGS-SLAM. Input-view PSNR uses SAM3 to hide dynamics, so it is not a novel-view score. The main paper does not report frame rate. Calling the result state of the art is fair on TUM and Bonn; the absolute lift over WildGS-SLAM is small.