Available at: https://digitalcommons.calpoly.edu/theses/3347
Date of Award
6-2026
Degree Name
MS in Computer Science
Department/Program
Computer Science
College
College of Engineering
Advisor
Jonathan Ventura
Advisor Department
Computer Science
Advisor College
College of Engineering
Abstract
Depth-supervised 3D Gaussian Splatting can improve scene geometry in indoor RGB-D capture, but noisy commodity depth measurements can also reduce rendered image quality. This thesis studies that trade-off in room-scale indoor reconstruction using two complementary depth sources: Kinect-class RGB-D sensor depth, which is metric but noisy and incomplete, and Depth Anything 3 (DA3) monocular depth, which is dense and meter-valued but not calibrated to the sensor or laser-reference geometry.
The proposed method constructs a per-pixel depth supervision target by inverse-variance fusion of the sensor depth and the monocular depth prediction. The fused target is used to train a depth-supervised 3D Gaussian Splatting model built on DN-Splatter. Two uncertainty models are evaluated: an implementation-based heuristic, using a configured Kinect noise floor and DA3 depth-structure terms, and a learned-confidence variant, using DA3's native reliability output. The method is evaluated on ten MuSHRoom indoor scenes under held-out Kinect depth metrics, with independent FARO point-cloud and per-frame depth protocols used for additional geometry validation.
Across the ten-room MuSHRoom evaluation, heuristic fusion improves RGB quality over sensor-only supervision in all 10 rooms, with mean improvements of 0.42 dB PSNR and 0.009 SSIM and a mean LPIPS reduction of 0.012. Kinect-evaluated geometry stays close to sensor-only supervision with a near-zero median AbsRel increase of 0.0009. FARO-based evaluation supports the same overall geometry conclusion while identifying one point-cloud completeness outlier. The DA3 learned-confidence variant gives mixed appearance changes but consistently worse geometry, showing that a foundation model's internal confidence is not automatically the right uncertainty model for metric fusion. These results support uncertainty-weighted sensor-plus-monocular depth fusion as a practical supervision strategy for indoor 3D Gaussian Splatting.