Radar-Anchored Metric Initialization
for 3D Gaussian Splatting under Suboptimal Illumination

RAMI teaser: task, idea, outcome
(a) Under poor illumination SfM triangulation collapses and random 3DGS becomes a basin lottery. (b) RAMI injects metric geometry once: a frozen monocular depth model is lifted to metric scale by a single weighted scale–shift fit to raw radar ranges, initializing an otherwise unchanged 3DGS pipeline. (c) The commodity radar matches the far more expensive LiDAR-cloud reference (24.03 vs. 24.19 dB) without using LiDAR at any stage.

Abstract

3D Gaussian Splatting (3DGS) is now the standard way for a mobile robot or a headset to synthesize novel views of a space it has walked through. Poor illumination destroys the structure-from-motion (SfM) points that bootstrap 3DGS, and supervising the model with range-sensor depth is not a substitute, because at single-chip angular resolution that loss helps or hurts unpredictably. We present RAMI, a radar-anchored metric initialization that injects sensor geometry at the one stage of 3DGS where it cannot conflict with photometric optimization. RAMI lifts relative depth from a frozen monocular foundation model to metric scale with one per-scene scale–shift fit against single-chip mmWave radar ranges, two scalars that vision cannot recover in the dark, and fuses the result into the initialization cloud of an unchanged 3DGS pipeline. Across an illumination-stratified indoor benchmark curated from a public 237-capture corpus, from bright rooms to lights-off, RAMI improves PSNR over matched random initialization by +1.11 dB on average and by up to +3.9 dB in the dark, and it matches a far more expensive LiDAR-cloud reference without using LiDAR at any stage. These results show that for mobile capture in the dark, a coarse range sensor is worth far more as a one-shot metric anchor than as a supervisor, which is what lets a commodity radar stand in for a LiDAR-grade initialization.

TL;DR. Poor light destroys the SfM points 3DGS starts from. A commodity single-chip radar is far too coarse to supervise a single pixel, yet it measures exactly the two scalars monocular depth is missing. Injected once as the initialization instead of attached as a loss, it places a whole scene at LiDAR-level quality.

Novel-View Rendering

Held-out novel views swept across each capture. Left to right: ground truth, randomly initialized 3DGS, and RAMI (same poses, same budget, same photometric loss; only the initial point set differs). Dark scenes are gamma-brightened for display.

Dark Lobby: lights-off capture (RAMI +3.88 dB over random)
Office (RAMI +2.67 dB)
Lounge (RAMI +1.72 dB)
Hall (RAMI +0.79 dB)

Drag to Compare

Drag the divider, or hover over it. Left of the handle is randomly initialized 3DGS, right is RAMI; same poses, same budget, same loss. Start with the crisp still frames to inspect detail, then switch to video to see the gain hold across viewpoints.

Dark Lobby +3.88 dB

Random Init  ↔  RAMI (Ours)
RandomRAMI RAMI Random
RandomRAMI
RAMI (Ours)  ↔  Ground Truth
RAMIGT GT RAMI
RAMIGT

Office +2.67 dB

Random Init  ↔  RAMI (Ours)
RandomRAMI RAMI Random
RandomRAMI
RAMI (Ours)  ↔  Ground Truth
RAMIGT GT RAMI
RAMIGT

Lounge +1.72 dB

Random Init  ↔  RAMI (Ours)
RandomRAMI RAMI Random
RandomRAMI
RAMI (Ours)  ↔  Ground Truth
RAMIGT GT RAMI
RAMIGT

Hall +0.79 dB

Random Init  ↔  RAMI (Ours)
RandomRAMI RAMI Random
RandomRAMI
RAMI (Ours)  ↔  Ground Truth
RAMIGT GT RAMI
RAMIGT

Stills are a single held-out view; videos sweep the held-out trajectory at 0.5× speed. Dark scenes are gamma-brightened for display.

Method

RAMI is deliberately minimal: no new losses, no trainable components, no change to the 3DGS optimizer. Only the initialization differs.

RAMI pipeline
(1) Inputs: low-light RGB, odometry poses, raw radar ADC (LiDAR is not used). (2) A frozen monocular model gives up-to-scale relative depth. (3) CFAR radar returns projected into views supply one weighted scale–shift fit per scene, which is the radar's sole role. (4) Metric depths are unprojected and voxel-fused into a 5k-point initialization cloud. (5) The cloud enters an otherwise unchanged 3DGS pipeline.

Qualitative Comparison

Eleven scenes (dark → bright) against recent pipelines. 3DGS-MCMC collapses on the dark tier, DN-Splatter renders are softer, and SfM initialization degrades where its triangulation yield collapses; RAMI stays sharp throughout.

Qualitative comparison across eleven scenes

Results

Comparison with recent pipelines on all 23 benchmark scenes (mean over scenes; strict training-frames-only protocol). Bold: best; underline: second best. RAMI leads the dark tier and the worst case, and stays within 0.16 dB of the matched-budget LiDAR-cloud reference on the overall mean.

MethodPSNR↑SSIM↑LPIPS↓ Dark↑Dim↑Bright↑Worst Δ / #fail
3DGS (random init)22.930.7570.549 25.8821.1723.86
+ SLV init (RAIN-GS-style)23.000.7590.554 25.7521.2324.06−2.49 / 11
+ per-view tone curves (Luminance-GS-style)20.610.7210.564 22.6819.6520.90−11.26 / 22
3DGS-MCMC16.020.4000.711 17.1415.9015.62−19.98 / 18
DN-Splatter-style (LiDAR loss)23.120.757 0.53026.3921.3723.89−1.58 / 10
3DGS + feed-forward init (DUSt3R/InstantSplat-style)23.140.7620.554 26.8820.9924.23−2.09 / 13
3DGS + SfM-points init23.960.7730.529 27.1522.1524.86−1.01 / 5
3DGS + LiDAR-cloud init24.190.776 0.52327.2322.3825.15−0.55 / 2
RAMI (ours, radar init)24.03 0.7740.52627.2622.27 24.84−0.22 / 2

Full 23-scene benchmark, ablations, training-seed/scale/anchor-sparsity controls, and the supervision-route negative results are in the paper and supplementary material (available upon acceptance).

BibTeX

@misc{rami2026,
  title = {Radar-Anchored Metric Initialization for 3D Gaussian
           Splatting under Suboptimal Illumination},
  year  = {2026}
}