MakeWorlds / Technical report
MAKEWORLDS TEAM · TECHNICAL REPORT
A system overview and selected experiments in image-based reconstruction across Mac, Linux and Windows.
Image-based reconstruction connects camera geometry, scene representation and image formation. This report describes the MakeWorlds workflow at the system level, from input photographs to a delivered Gaussian scene. It introduces the public concepts behind the representation and explains how visual comparisons and numerical evaluation answer different questions. Experiments on Train, Dr Johnson and Playroom report image quality and cross-platform variation for a fixed candidate, with the published 3DGS values as reference points.
Topics Image-based reconstruction · Gaussian splatting · Camera estimation · Cross-platform evaluation
01 / SYSTEM OVERVIEW
The task is to recover a scene that can explain photographs taken from different positions. A useful result depends on both the estimated cameras and the learned scene: their relationship determines how a three-dimensional model becomes an image.
Multiple observations of one scene
Relate images to a shared scene
Fit a Gaussian representation
Export the Gaussian PLY result
Compare rendered and reference views
Camera estimation establishes the viewpoint associated with each image. Intrinsics describe the camera projection, while extrinsics relate the camera to scene coordinates. Structure-from-Motion provides this geometric context; the COLMAP documentation describes the underlying reconstruction problem.
MakeWorlds carries the reconstruction through to a Gaussian PLY asset. The visual study pairs images from the delivered scene with their reference photographs. This gives readers an observable result alongside the numerical summary.
02 / TECHNICAL BACKGROUND
A Gaussian scene uses spatial primitives with position, oriented extent, opacity and appearance. These primitives exist in a shared three-dimensional coordinate system and can be projected into a chosen camera. This representation is established in the 3D Gaussian Splatting paper.
During scene fitting, rendered views are compared with observed images. Changing the scene parameters changes the rendered images, allowing image disagreement to guide optimization. The expression alongside gives the standard spatial Gaussian used to describe the primitive in that literature.
Spatial Gaussian · public formulation
g(x) = exp[−½(x − μ)ᵀ Σ⁻¹(x − μ)]
03 / QUALITATIVE STUDY
Tanks and Temples · 301 photographs
A locomotive reconstructed from photographs. Compare the delivered scene with the original photograph at the same test viewpoint.
MakeWorlds render
Original photograph04 / EXPERIMENTAL DESIGN
Camera estimation, scene quality and platform variation address different parts of the reconstruction problem.
Each image pair uses one evaluated camera viewpoint. The display views highlight overall shape, an oblique overview and side texture. Numerical summaries instead aggregate the evaluated test views, so they do not depend on which images appear in the visual study.
The paper-comparison table places the recorded PSNR, SSIM and LPIPS beside published per-scene reference values. MakeWorlds estimates its own cameras, so these figures describe an end-to-end reconstruction comparison. The separate cross-platform record asks how far the test-set means vary across the three platform runs.
The cross-platform spread is computed within its own evaluation protocol. It should not be reconstructed by subtracting values from the paper-comparison table. Keeping the protocols separate preserves what each measurement actually describes.
bb039199Image-quality metrics characterize this set of reconstructions. Platform spread characterizes agreement of the measured image-quality means; it is not a frame-rate measurement or a claim of byte-identical models. The data files retain the candidate and run identifiers so each row can be traced to its recorded evaluation.
05 / EXPERIMENTAL RESULTS
Three scene studies, each measured on three platforms. PSNR and SSIM describe image similarity; LPIPS measures perceptual difference.
| Scene | Source / platform | PSNR ↑ dB | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|---|
| Train | Paper reference | 21.097 | 0.8020 | 0.2180 |
| Train | Mac | 21.227 | 0.8463 | 0.1537 |
| Train | Linux | 21.282 | 0.8479 | 0.1528 |
| Train | Windows | 21.205 | 0.8468 | 0.1530 |
| Dr Johnson | Paper reference | 28.766 | 0.8990 | 0.2440 |
| Dr Johnson | Mac | 29.180 | 0.9124 | 0.2215 |
| Dr Johnson | Linux | 29.170 | 0.9101 | 0.2267 |
| Dr Johnson | Windows | 29.274 | 0.9095 | 0.2271 |
| Playroom | Paper reference | 30.044 | 0.9060 | 0.2410 |
| Playroom | Mac | 30.385 | 0.9148 | 0.2264 |
| Playroom | Linux | 30.435 | 0.9158 | 0.2252 |
| Playroom | Windows | 30.205 | 0.9137 | 0.2268 |
Higher PSNR / SSIM and lower LPIPS are better. Paper rows are published reference values; MakeWorlds solves its own cameras. These are selected scene results, not an exhaustive benchmark.
Maximum minus minimum PSNR, using the separate cross-platform evaluation protocol.
0.072 dB
0.104 dB
0.230 dB
06 / REFERENCES
Public literature for the underlying concepts, and recorded inputs for the experiments.