Stable Single-Pixel Contrastive Learning for Semantic and Geometric Tasks

TL;DR

This paper introduces a stable contrastive loss for pixel-level representations that are view-invariant and semantically meaningful, enabling precise correspondence without momentum-based training, demonstrated in synthetic environments.

Contribution

It proposes a novel stable contrastive loss for pixel representations that captures semantic and geometric info without momentum-based training.

Findings

Effective pixel correspondence in synthetic environments

View-invariant and semantically meaningful descriptors

Stable contrastive loss outperforms prior methods

Abstract

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.