about fault detection marathon

got prize 7th place on fault detection marathon on topcoder.
the task was to reconstruct and cluster curved mesh surfaces from probability volume
my approach was
1) threshold and subsample points by poisson resampling
2) reconstruct mesh by ball-pivoting algorithm, this gave connectivity and normals
3) generate for each mesh point its connectivity feature vector
via approach from paper "Multiple Manifold Clustering Using Curvature Constrained Path",
the modifications were that curvature constraint I measured via edge normals,
I also tried to extrapolate by plane to pass through intersection areas.
4) hdbscan clustering