Point geometry
Positions provide scene structure and can seed Gaussian locations. Density, noise and coordinate alignment affect the initialization.
Conversion guide
Point Cloud to Gaussian Splat Conversion
A point cloud can provide strong geometric initialization, but geometry alone is not the same as a view-synthesis dataset. Useful conversion depends on RGB observations, camera poses, scan alignment and scene coverage.
Feasibility
When conversion is possible
The strongest source package combines measured geometry with calibrated appearance observations.
RGB information
Per-point color can help, but projected color alone may not contain enough multi-view appearance to render novel viewpoints accurately.
Source imagery
Overlapping photos or panoramas provide the observations used to optimize appearance across viewpoints. Consistent exposure and sharpness help.
Camera poses
Camera intrinsics and positions/orientations connect each image to the scene. Missing or inaccurate poses make training unreliable.
Containers
E57, LAS, LAZ and PLY are not equivalent inputs
| Format | Common contents | e57.gs automated status |
|---|---|---|
| E57 | Aligned scan positions, points, color and sometimes imagery or pose-related data | Accepted by the automated account upload |
| LAS / LAZ | LiDAR points, classification, intensity and optional RGB; LAZ is compressed | Not advertised as a direct automated input |
| PLY | Flexible point, mesh or Gaussian attributes; schema varies widely | Not advertised as a direct automated input |
Inside E57: E57 is a container rather than a simple XYZ list. It stores measured point data in data3D and may also include color, embedded imagery, and camera or pose-related information in images2D. Export contents vary, so the file extension alone does not guarantee that usable imagery or camera information is present.
For LAS, LAZ or point-cloud PLY, prepare a supported E57 package with scan positions aligned in one coordinate system where appropriate, or request a feasibility review. Renaming an extension does not convert the underlying data.
Output quality
Initialization is only the beginning
Gaussian optimization can refine a scene representation, but it cannot create trustworthy detail for surfaces and viewpoints absent from the capture.
- Dense, correctly aligned geometry can stabilize structure but does not replace image observations.
- Sparse viewpoints create holes, floaters or unstable appearance at novel angles.
- Reflective, transparent or moving surfaces can violate the assumptions of a static scene.
- A completed result must be visually inspected; it is not certified survey or engineering data.
Deliverables
Choose output by viewer and tool support
Gaussian PLY
Useful for compatible Gaussian-Splat editors and interchange workflows. A normal point-cloud PLY and a Gaussian PLY are not interchangeable merely because both use .ply.
SPLAT
A compact distribution format used by compatible web viewers. It may omit attributes needed by editing tools, so keep the richer output when available.
Working with a Matterport Pro3 export?
Use the dedicated workflow for a Matterport E57 and see what the converter checks before processing.
Matterport Pro3 E57 to Gaussian SplatFAQ
Point-cloud conversion questions
Can every point cloud become a Gaussian Splat?
No. Geometry alone can initialize a scene, but high-quality view synthesis also needs suitable color observations, coverage and camera information. Sparse, uncolored or noisy scans have stricter limits.
Does e57.gs accept LAS, LAZ or PLY directly?
The automated e57.gs upload currently accepts E57. LAS, LAZ and PLY are not advertised as direct automated inputs; they may need preparation into a supported E57 package or a feasibility review.
Why do camera poses matter?
Camera poses connect each image observation to a viewpoint in 3D space. Without reliable poses, a training pipeline cannot consistently assign appearance to the scene from multiple views.
What limits output quality?
Missing viewpoints, occlusion, motion, reflective surfaces, low-resolution color, scan-alignment errors, noise and large changes in exposure can all create gaps or artifacts.
Start with a supported E57
The automated service accepts E57. Check that your export includes aligned point geometry and the best available color or image data.
Explore E57 to Gaussian Splat