EliGSiR: Continual RGB-D Mapping with Gaussian Splatting under Bounded Compute
Björn Ellensohn, Elmar Rueckert, Christian Rauch
cs.CV, cs.RO
2026-09-17
EliGSiR retargets views, resolution, and geometry growth under a compute budget. On TUM it hits 21.52 dB vs SplaTAM 19.42 dB, in 176 s vs 1383 s.
Offline 3D Gaussian Splatting assumes a closed observation set and a long schedule. Continual RGB-D mapping keeps receiving new views while old regions must stay intact, under a compute cap. Time spent on a redundant view is time not spent on a new wall. Full-resolution steps are expensive, so fewer of them fit. Growing the Gaussian population makes every later step heavier.
EliGSiR writes continual Gaussian mapping as three runtime decisions: which views to optimize, at what image fidelity, and where to add geometric capacity. Poses can come from ground truth, visual odometry, or ORB-SLAM3; the system is a mapper, not a full tracker.
Map-guided view scheduling pose-filters redundant frames, then admits views by overlap with the current Gaussian map. Admission is not a lifetime keyframe: a view can drop in priority once its region is explained and rise again if error returns. The scene is partitioned into fixed world-space microcells; reconstruction error already computed during training renders raises replay priority on weak cells.
Load-adaptive fidelity treats supervision resolution as a reversible switch. When recent mapping cost is high, resolution steps down through 1/8, 1/4, and 1/2; when the queue drains, it steps back up. Source RGB-D is kept, intrinsics scale with the grid, so fine detail remains available later.
Targeted geometry growth splits depth supervision from Gaussian creation. Newly seen space first gets a bounded set of sparse depth-derived Gaussians. Extra primitives appear only where repeated RGB-D views still disagree with the map. Population updates reuse 3DGS-MCMC relocation, replacement, and densification; that mechanism is not claimed as a contribution.
Evaluation uses Replica, TUM RGB-D, ScanNet++, and real Orbbec sequences. The main table shares ground-truth mapping poses. CaRtGS cannot take external poses, so that row compares EliGSiR with live ORB-SLAM3. Besides final PSNR, the paper reports causal visible quality (CVQ) during acquisition, its area under the source-progress curve, and the fraction of measurable cells above 20 dB (CUC@20).
Final quality and wall time have to be read together.
| Sequence | Method | PSNR | Time |
| TUM fr3/office, GT poses | EliGSiR | 21.52 dB | 176.1 s |
| same | SplaTAM | 19.42 dB | 1382.6 s |
| same | RTG-SLAM | 14.46 dB | 1147.8 s |
| TUM, tracked poses | EliGSiR + ORB-SLAM3 | 23.02 dB | 155.5 s |
| same | CaRtGS native tracker | 20.10 dB | 230.9 s |
| Orbbec kitchen1 | EliGSiR | 23.50 dB | 203.3 s |
| Replica office0 | EliGSiR | 34.20 dB | 324.0 s |
| same | VarSplat +30k steps | 40.91 dB | 3742.2 s |
The TUM sequence lasts 87.14 s. EliGSiR’s tracked run is 1.78× real time, CaRtGS 2.65×, and the GT-pose baselines 13–29×. During acquisition the full method’s CVQ-AUC is 21.18 dB versus 20.62 without adaptive fidelity and 20.84 without targeted growth. On TUM, adaptive fidelity cuts supervised pixels from 2786.9 M to 398.2 M, about 7×, and mapping time from 103.34 s to 89.15 s; fixed native resolution still wins final PSNR (19.36 vs 18.50) and loses online quality. On ScanNet++, targeted growth shrinks the map from 1.109 M to 366.5 k Gaussians with almost the same endpoint PSNR. View scheduling does not lift the mean; several weak cells move by more than 10 dB. After the stream, 16k extra updates take PSNR from 23.27 to 25.62 dB.
On the real floor2 sequence EliGSiR ends at 16.27 dB, below SplaTAM at 21.40 and VarSplat at 25.88, with a much shorter clock. Budget versus quality, not a sweep of first place.
Most online Gaussian SLAM papers report a final PSNR. This one scores whether the map is usable while the robot is still moving, and treats resolution, replay, and densification as a shared budget. For anyone who needs the map before the loop ends, 21.5 dB in 176 s is closer to deployable than 19.4 dB in 1383 s. The code sits on gsplat, runs on NVIDIA and AMD, and ships a ROS 2 package.
The main setting assumes a mostly static scene and a stable trajectory. Pose error, movers, and bad depth still break geometry. Real-time needs hand-tuned runtime knobs, and the representation grows with explored space. This is a mapper; a full SLAM system still needs a visual odometry frontend. On floor2 the final color score lags. Adaptive fidelity on kitchen1 trades endpoint PSNR from 22.91 down to 18.25 for better in-progress metrics. Experiment blocks use different splits, so absolute PSNR must not be compared across tables.