Why Can 2,000 Photos Still Produce a Broken 3D Gaussian Splat?
Image count is not geometric evidence. Diagnose coverage, parallax and camera poses before adding more photos.
Research note ·
VISULOMA JOURNAL
Recommended for you
Real-Space Digitization Research
VISULOMA RESEARCHAre 3DGS Floaters Caused by COLMAP? Camera Pose Errors and Their Limits
Low reprojection error is not proof of correct surfaces. Separate calibration, matching, pose and training failures.
Research note ·
Why Is a 3D Gaussian Splat Sharp from the Front but Broken from the Side?
Separate extrapolation, occlusion and view-dependent appearance, then evaluate the intended viewing envelope.
Research note ·
Can AI Reconstruct a Real Home from 10 Photos? Sparse Views versus Generated Geometry
Sparse inputs can produce plausible views. Unobserved rooms still require evidence and validation.
Research note ·
Can AI Masks Remove Moving People from a 3DGS Capture?
Segmentation is only one step. Feature extraction, training losses and hidden background need separate checks.
Research note ·
Can LiDAR Reduce 3DGS Reconstruction Failures? Geometry, Poses and Depth
LiDAR adds geometry evidence, not a universal appearance fix. Evaluate calibration, synchronization and training integration.
Research note ·
Postshot vs Nerfstudio: How to Compare 3DGS Quality Fairly
Separate usability, experimental control and rendering quality. Compare the same evidence and delivery conditions.
Research note ·
3D Gaussian Splatting vs Mesh for Web-Based Real-Space Experiences
Choose by appearance, editability, collisions and loading constraints, then validate the target experience.
Research note ·