Learning Location-Aware Visual Representations Through Anchor-Based Contrastive Learning

Chandan Kumar

doi:10.1007/978-3-032-07109-5_18

Return

Learning Location-Aware Visual Representations Through Anchor-Based Contrastive Learning

2026-01-01

PRE

Chandan Kumar *

Jansel Herrera-Gerena

John Just

Matthew J. Darr

Ali Jannesari

DOI：10.1007/978-3-032-07109-5_18

Original

Original request for help

Save

Abstract

En 中文

Visual representation learning has made significant strides, yet existing approaches often overlook the crucial aspect of spatial understanding within images. While current self-supervised methods excel at learning semantic features, they typically do not explicitly encode location information, limiting their utility for downstream tasks that require spatial reasoning. We present a novel approach that learns location-aware visual representations through contrastive learning without requiring labeled data. Our method introduces an anchor-based strategy that explicitly encourages the model to capture both semantic and spatial information in its learned representations. By pioneering the integration of intra-image contrastive learning alongside traditional inter-image comparisons, our approach enables the extraction of rich spatial features while maintaining semantic understanding. We demonstrate the effectiveness of our learned representations through comprehensive similarity analysis, achieving 89.2% accuracy in spatial correspondence tasks - a 15-fold improvement over randomly initialized features. Our method opens new possibilities for self-supervised learning by showing how contrastive approaches can be extended to capture spatial relationships while maintaining semantic understanding.

Keywords:

Unsupervised learning

Representation learning

Object detection