[Paper Review] Hierarchical Invariant Feature Learning with Marginalization for Person Re-Identification

Rahul Rama Varior, Gang Wang|arXiv (Cornell University)|Nov 30, 2015

Video Surveillance and Tracking Methods41 references3 citations

TL;DR

This paper proposes a hierarchical invariant feature learning framework with marginalization for person re-identification, leveraging labeled image pairs to learn pose- and illumination-invariant features in a kernel space. By integrating marginalization into SVM metric learning, it achieves state-of-the-art performance on VIPeR, CUHK01, CAVIAR4REID, and iLIDS without explicit data augmentation, improving generalization on small datasets.

ABSTRACT

This paper addresses the problem of matching pedestrians across multiple camera views, known as person re-identification. Variations in lighting conditions, environment and pose changes across camera views make re-identification a challenging problem. Previous methods address these challenges by designing specific features or by learning a distance function. We propose a hierarchical feature learning framework that learns invariant representations from labeled image pairs. A mapping is learned such that the extracted features are invariant for images belonging to same individual across views. To learn robust representations and to achieve better generalization to unseen data, the system has to be trained with a large amount of data. Critically, most of the person re-identification datasets are small. Manually augmenting the dataset by partial corruption of input data introduces additional computational burden as it requires several training epochs to converge. We propose a hierarchical network which incorporates a marginalization technique that can reap the benefits of training on large datasets without explicit augmentation. We compare our approach with several baseline algorithms as well as popular linear and non-linear metric learning algorithms and demonstrate improved performance on challenging publicly available datasets, VIPeR, CUHK01, CAVIAR4REID and iLIDS. Our approach also achieves the stateof-the-art results on these datasets.

Motivation & Objective

• Address the challenge of person re-identification under varying lighting, pose, and environmental conditions.
• Overcome the limitation of small labeled datasets in person re-identification by enabling effective generalization from limited data.
• Develop a feature learning framework that captures invariant representations across camera views using labeled image pairs.
• Integrate marginalization into metric learning to simulate training on large, augmented datasets without explicit data corruption.
• Achieve state-of-the-art performance on benchmark datasets with improved robustness and generalization.

Proposed method

• Extract local features from images and map them into a kernel space using a non-linear kernel function.
• Learn a linear transformation in the kernel space to enforce invariance for corresponding parts of matching image pairs.
• Form a global image representation by concatenating the marginalized mapped local features.
• Adapt SVM Metric Learning (SVMML) by incorporating marginalization to handle noisy or corrupted inputs implicitly.
• Use the marginalized features as input to a metric learning framework to improve classification performance on test data.
• Train the hierarchical network end-to-end using labeled image pairs to optimize for inter-personal similarity and intra-personal invariance.

Experimental results

Research questions

• RQ1 Can a hierarchical feature learning framework effectively learn invariant representations across camera views for person re-identification?
• RQ2 How can marginalization in metric learning improve generalization when training data is limited?
• RQ3 Does combining kernel-based feature mapping with marginalized metric learning outperform existing state-of-the-art methods on standard benchmarks?
• RQ4 To what extent does the proposed framework reduce sensitivity to pose, lighting, and environmental variations?
• RQ5 Can the marginalization technique simulate the benefits of data augmentation without increasing computational cost?

Key findings

• The proposed method achieves a Rank-1 accuracy of 47.9% on the VIPeR dataset, surpassing all prior state-of-the-art methods.
• On the CAVIAR4REID dataset, the method achieves 45.1% Rank-1 accuracy when combined with kLFDA, outperforming all other methods except MFA.
• On the iLIDS dataset, the method achieves 47.7% Rank-1 accuracy when combined with kLFDA, ranking second at higher ranks.
• On the CUHK01 dataset, the method achieves 39.5% Rank-1 accuracy when combined with kLFDA, outperforming all baselines and existing methods.
• The marginalization technique significantly improves generalization, especially on small datasets, by implicitly simulating data augmentation.
• The framework demonstrates strong robustness to pose and illumination variations, as evidenced by consistent performance gains across all benchmark datasets.