Computer vision · Research reproduction

EDGS reproduction

An independent reproduction of 3D Gaussian initialisation, validated against analytic scene geometry.

My roleImplemented the NumPy pipeline, independent solver checks, and a parallax guard for poorly conditioned geometry.

Research reproduction · Initialisation validated

Explore the workflow

Interactive walkthrough

A good match can still have poor geometry.

Compare stable and near-parallel viewing rays, then test the optional parallax guard.

Illustrative viewing-ray pair

Add the optional 3° minimum-angle guard

Select a stage to inspect its role in the workflow.

Step 4: Initialise. Initialise a Gaussian from the surviving point.

Two-view geometry / IllustrativePoint retained

Recorded stress-scene result0.0486median surface error · scene units

Paper filters, no angle guardSeparate synthetic experiment · rounded from raw data

Step 04 / Inspect

Initialise a Gaussian from the surviving point.

Accepted points receive colour coefficients, scale, opacity, and rotation before export to a standard 3DGS PLY. This example illustrates the decision rather than running that pipeline.

point: retained
output: 3DGS point-cloud PLY
full training: outside this reproduction

6° illustrative pair. Accepted for initialisation.

Illustrative geometry, not to scale. Measurements come from the separate synthetic stress scene.

Inspect the synthetic results

Results

84 tests covering camera geometry, triangulation, sampling, spherical harmonics, input/output, and pipeline integration.
Noiseless synthetic recovery to numerical precision; independent solver agreement within 1e-8.
On the synthetic stress scene, a 3° parallax guard reduced median surface error from 0.0486 to 0.0042 scene units.

Engineering details

Near-parallel rays can pass confidence and reprojection filters while producing inaccurate 3D points.

Reproduce the pipeline

Reimplement correspondence alignment, triangulation, filtering, spherical-harmonics fitting, and standard 3DGS PLY export.

Check known geometry

Measure error against analytic surfaces and cross-check triangulation with an independent SVD solver.

Filter unstable pairs

An optional minimum-angle guard rejects poorly conditioned rays; it remains disabled in the paper-faithful configuration.

More work