2512.04970

Total: 1

#1 Stable Single-Pixel Contrastive Learning for Semantic and Geometric Tasks [PDF1] [Copy] [Kimi] [REL]

Authors: Leonid Pogorelyuk, Niels Bracher, Aaron Verkleeren, Lars Kühmichel, Stefan T. Radev

We pilot a family of stable contrastive losses for learning pixel-level representations that jointly capture semantic and geometric information. Our approach maps each pixel of an image to an overcomplete descriptor that is both view-invariant and semantically meaningful. It enables precise point-correspondence across images without requiring momentum-based teacher-student training. Two experiments in synthetic 2D and 3D environments demonstrate the properties of our loss and the resulting overcomplete representations.

Subject: Computer Vision and Pattern Recognition

Publish: 2025-12-04 16:38:26 UTC