title: How to fix RANSAC global registration that fails (FPFH features, voxel size, mutual filter) description: Why FPFH + RANSAC global registration returns a low fitness or a wrong pose on real scans: voxel size, feature radius, units, symmetry, the mutual filter; what pardesline sets and how to diagnose with ransac_fitness.

How to fix RANSAC global registration that fails¶

The symptom¶

ransac_fitness is 0.05 and the pose is random; or RANSAC takes a minute and returns a pose that ICP then "refines" into a wrong minimum.

Causes, in the order to check them¶

Units. A millimetre cloud with a voxel chosen for metres has one point per cell. obj.frame.units, then with_frame(units="mm").
Voxel size. Features are computed on the downsampled cloud. Too small: the FPFH histograms describe noise. Too large: there is nothing left to describe. Rule: 2–5× the point spacing, about 1 % of the bounding-box diagonal. align() picks 1.3 % of the diagonal when voxel_size=0.
Feature radius. FPFH needs about five voxels of support (fpfh_radius_factor=5); normals need two (normal_radius_factor=2). Larger radii blur partial views, smaller ones make every point look the same.
Mutual filter off. Without it, one-directional matches flood RANSAC with false correspondences; mutual_filter=True is the single most effective setting on symmetric or repetitive parts.
Overlap below 30 %. FPFH cannot match what is not there. Crop both clouds to the region that overlaps, or acquire more.
Symmetry. Cylinders, spheres, planes: several poses score the same. ransac_n=4, a prior from the fixture, or landmarks.

What pardesline sets¶

import pardesline as pl

coarse = target.run("registration.ransac_global", moving=moving)  # inspect the global stage alone
print(coarse.transform["fitness"], coarse.transform["inlier_rmse"])
print(pl.describe("registration.ransac_global"))  # every parameter with its rationale

voxel_size=0 (auto), normal_radius_factor=2, fpfh_radius_factor=5, distance_threshold_factor=1.5, mutual_filter=True, ransac_n=3, edge_length_threshold=0.9, ransac_max_iteration=100000, ransac_confidence=0.999. The rationale for each is on the registration page.

A quick experiment¶

for v in (0.5, 1, 2, 4):  # multiples of the auto voxel, in file units
    r = target.run("registration.ransac_global", moving=moving, voxel_size=v * 0.013 * diag)
    print(v, r.transform["fitness"], r.transform["inlier_rmse"])

Pick the voxel with the highest fitness at a sane RMSE, then run the full pipeline with it.