Adaptive sparse contrastive learning for unsupervised object re-identification

doi:10.1016/j.patcog.2025.112604

Adaptive sparse contrastive learning for unsupervised object re-identification

2025-10-15

0

PRE

AI

DOI：10.1016/j.patcog.2025.112604

Original

Original request for help

Save

Abstract

En 中文

Our contributions can be summarized in three parts: • We propose a novel contrastive learning paradigm called adaptive sparse contrastive learning (ASCL), which is different from the conventional dense contrastive learning mechanism. This is the first time such a paradigm has been applied to unsupervised object re-ID. ASCL allows the model to focus on a select few contrastive pairs that contain reliable identity cues, thereby guiding the learning of more discriminative feature representations. • We develop two complementary sparse contrastive learning objectives that can adaptively adjust the within-class pulling strength of positive contrastive pairs according to the dynamically changed intra-cluster discrepancy. These two objectives collaboratively introduce more effective contrastive signals for each pseudo-class, thereby mitigating the efficiency-reliability trade-off commonly encountered in unsupervised learning. • We conducted extensive experiments on three large-scale datasets for object re-ID. The results validate the effectiveness of our proposed approach, demonstrating that it significantly outperforms state-of-the-art methods across all three datasets. Specifically, our method achieves 46.6 % mAP and 76.4 % top-1 accuracy on the challenging MSMT17 dataset.

Adaptive sparse contrastive learning for unsupervised object re-identification

Abstract

Journal

Organization