[Paper Review] Highly Efficient Regression for Scalable Person Re-Identification

Hanxiao Wang, Shaogang Gong+1 more|arXiv (Cornell University)|Dec 5, 2016

Video Surveillance and Tracking Methods3 citations

TL;DR

This paper proposes Highly Efficient Regression (HER) for scalable person re-identification, framing re-ID as a ridge regression problem enhanced with Fisher’s criterion to achieve fast, accurate learning. HER+ enables real-time incremental updates, outperforming state-of-the-art models in accuracy and reducing human labeling effort by 50% with active learning, while being up to 89.7× faster than batch retraining on large datasets.

ABSTRACT

Existing person re-identification models are poor for scaling up to large data required in real-world applications due to: (1) Complexity: They employ complex models for optimal performance resulting in high computational cost for training at a large scale; (2) Inadaptability: Once trained, they are unsuitable for incremental update to incorporate any new data available. This work proposes a truly scalable solution to re-id by addressing both problems. Specifically, a Highly Efficient Regression (HER) model is formulated by embedding the Fisher's criterion to a ridge regression model for very fast re-id model learning with scalable memory/storage usage. Importantly, this new HER model supports faster than real-time incremental model updates therefore making real-time active learning feasible in re-id with human-in-the-loop. Extensive experiments show that such a simple and fast model not only outperforms notably the state-of-the-art re-id methods, but also is more scalable to large data with additional benefits to active learning for reducing human labelling effort in re-id deployment.

Motivation & Objective

• Address the scalability limitations of existing person re-identification (re-ID) models, which suffer from high computational cost and lack of incremental learning support.
• Develop a model with low complexity and scalable memory usage to enable training on large-scale, cumulative datasets without retraining from scratch.
• Enable real-time, incremental model updates to support human-in-the-loop active learning, reducing reliance on extensive pre-labeled data.
• Improve re-ID performance while maintaining computational efficiency, making the model suitable for real-world deployment.

Proposed method

• Formulate person re-identification as a multivariate ridge regression problem, using a closed-form solution for fast training and inference.
• Integrate Fisher’s criterion into the ridge regression objective to maximize class separability and preserve discriminative power without sacrificing speed.
• Design HER+, an incremental extension of HER that supports online model updates by efficiently updating the regression weights upon arrival of new labeled data.
• Implement a joint exploration-exploitation active learning strategy (Joint E²) to guide human labeling toward the most informative samples, minimizing labeling cost.
• Use LSQR for efficient large-scale computation, enabling scalability to datasets with over 10⁴ samples.
• Apply the model to both batch and incremental learning settings, comparing performance and training time against state-of-the-art re-ID methods.

Experimental results

Research questions

• RQ1 Can a simple regression-based model achieve state-of-the-art re-ID performance while being significantly faster and more memory-efficient than existing complex models?
• RQ2 Can Fisher’s criterion be effectively embedded into ridge regression to maintain discriminative power without increasing computational cost?
• RQ3 Can incremental learning be achieved in re-ID with real-time model updates that avoid full retraining?
• RQ4 Can active learning guided by HER+ significantly reduce human labeling effort while maintaining or improving re-ID accuracy?
• RQ5 How does the computational cost of incremental updates with HER+ compare to full retraining on large-scale re-ID datasets?

Key findings

• HER achieves state-of-the-art re-ID performance on VIPeR and CUHK01, with Rank-1 accuracy of 45.2% and 64.7% respectively when trained on 100% labeled data.
• HER+ achieves 55.8% Rank-1 accuracy on CUHK01 with only 50% labeled data using active learning, surpassing models trained on 100% data using random labeling.
• Active learning with the Joint E² criterion reduces labeling effort by 50% while achieving higher accuracy than random labeling at the same budget.
• HER+ performs incremental updates in 0.03s, 0.2s, and 0.6s on VIPeR, CUHK01, and CUHK03, respectively—11.5 to 89.7× faster than batch retraining.
• HER+ is over 10,000× faster than retraining conventional models like kLFDA, XQDA, and MLAPG when scaling to large datasets.
• Incremental updates with HER+ maintain model performance comparable to full batch retraining, confirming no accuracy degradation from incremental learning.