Stable Single-Pixel Contrastive Learning for Semantic and Geometric Tasks

2025-12-04

AI Summary

This work addresses the challenge of jointly modeling semantic and geometric information in pixel-level representation learning, aiming to achieve precise cross-image point correspondence without momentum-based mechanisms. We propose a novel stable contrastive loss that eliminates the conventional momentum teacher-student architecture and enables, for the first time, end-to-end, single-pixel-level joint semantic-geometric representation learning. Our method leverages overcomplete feature maps and pixel-wise contrastive learning, trained self-supervisedly on synthetic 2D/3D environments. Experimental results demonstrate that the learned representations exhibit both strong semantic discriminability and high geometric fidelity, leading to significant improvements in cross-view point matching accuracy. This approach establishes a new paradigm for unsupervised pixel-level alignment, advancing beyond reliance on momentum-based consistency or hand-crafted geometric priors.

Technology Category

• Computer Vision: Representation Learning for Vision
• Machine Learning: Representation Learning
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

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.

Problem

• Develop stable contrastive losses for pixel-level representation learning
• Create view-invariant and semantically meaningful overcomplete descriptors
• Enable precise point-correspondence without momentum-based teacher-student training

Innovation

• Stable contrastive losses for pixel-level representations
• Overcomplete descriptors for view-invariant semantics
• Point-correspondence without momentum-based teacher training