NeurIPS 2026 · Evaluations & Datasets Track · Spotlight

What Downstream Scores Miss in Point-Cloud Self-Supervised Learning

Kohsuke Ide · Ryousuke Yamada · Yue Qiu · Yoshihiro Fukuhara · Hirokatsu Kataoka · Yuki M. Asano · Yutaka Satoh

Overview of controlled diagnostics for point-cloud self-supervised learning.

Overview

Downstream scores measure performance, but do not by themselves explain the source of a gain. We study controlled diagnostics for the pretext dependency, the visible geometric support, and the downstream readout in point-cloud self-supervised learning.

Paper and code links will be added when available.