GEKey: Learning Geodesic Eccentric 3D Keypoints via Self-Supervision

Pierre Onghena, Santiago Velasco-Forero, Beatriz Marcotegui

Mines Paris – PSL University Centre for Statistics and Images (STIM)

BMVC Oral 2026

Four stages on the Stanford dragon: point cloud, geodesic eccentricity transform, local maxima, interest points.

The geodesic eccentricity function reveals a set of salient, object-specific points from the 3D point cloud to guide our self-supervised method.

54.5% Detection mIoU
73.0% Alignment DAS
Stable Under non-rigid deformation

Abstract

Point clouds represent 3D object geometry as sets of spatial coordinates. A fundamental task in their analysis is keypoint detection, i.e., identifying a sparse subset of points that captures the most salient features. While supervised methods rely on manual annotations, unsupervised approaches leverage intrinsic geometric properties to localize keypoints without human input. We introduce GEKey, a self-supervised framework that uses the geodesic eccentricity transform, an SE(3)-invariant geometric measure, to identify instance-specific salient points. These points serve as pseudo-labels to train a model that aligns its predictions with the object’s intrinsic structure. GEKey is simple, efficient, and robust, preserving intra-class variation by balancing instance-specific and category-level structure. Extensive experiments on ShapeNetCoreV2, KeypointNet, CAPE, ITOP, and DeformingThings4D show state-of-the-art performance in both keypoint detection and alignment, with strong resilience to noise, downsampling, deformations, and real-world conditions.

Method Overview

Pipeline: a backbone and MLP predict keypoints, compared through a bipartite matching loss to pseudo-labels from the geometric analysis.
Overview of self-supervised GEKey to generalize keypoint prediction. The blue workflow shows a backbone model predicting keypoints Ŷ, which are compared, via the bipartite matching loss, to pseudo-labels Y* generated by the green pipeline.

Citation

@inproceedings{onghena2026gekey,
    title     = {{GEKey}: Learning Geodesic Eccentric {3D} Keypoints via Self-Supervision},
    author    = {Onghena, Pierre and Velasco-Forero, Santiago and Marcotegui, Beatriz},
    booktitle = {British Machine Vision Conference (BMVC)},
    year      = {2026}
}