Person Re-Identification by Dual-Regularized KISS Metric Learning

Abstract

Person re-identification aims to match the images of pedestrians across different camera views from different locations. This is a challenging intelligent video surveillance problem that remains an active area of research due to the need for performance improvement. Person re-identification involves two main steps: feature representation and metric learning. Although the keep it simple and straightforward (KISS) metric learning method for discriminative distance metric learning has been shown to be effective for the person re-identification, the estimation of the inverse of a covariance matrix is unstable and indeed may not exist when the training set is small, resulting in poor performance. Here, we present dual-regularized KISS (DR-KISS) metric learning. By regularizing the two covariance matrices, DR-KISS improves on KISS by reducing overestimation of large eigenvalues of the two estimated covariance matrices and, in doing so, guarantees that the covariance matrix is irreversible. Furthermore, we provide theoretical analyses for supporting the motivations. Specifically, we first prove why the regularization is necessary. Then, we prove that the proposed method is robust for generalization. We conduct extensive experiments on three challenging person re-identification datasets, VIPeR, GRID, and CUHK 01, and show that DR-KISS achieves new state-of-the-art performance.

Cross-Dataset Person Re-identification Using Similarity Preserved Generative Adversarial Networks

“…Supervised Learning: Most existing person Re-ID models are supervised, and based on either invariant feature learning [8] , metric learning [12] or deep learning [16] . However, in the practical deployment of Re-ID algorithms in large-scale camera networks, it is usually costly and unpractical to label the massive online surveillance videos to support supervised learning as mentioned in [11] .…”

“…Due to the privacy problem regarding the collection of surveillance videos and the expensive cost of data labeling, most of the proposed Re-ID algorithms [16] [12] conduct supervised learning on small labeled datasets. Directly deploying these trained models to the real-world large-scale camera networks may lead to a poor performance, because the images captured from different camera networks usually have totally different backgrounds, noise distributions, brightness and resolution as shown in Fig.1.…”

Triplet-Based Deep Hashing Network for Cross-Modal Retrieval

“…Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. graph regularization [6] [7], and dictionary learning [8], are based on subspace learning [9]- [11], which maps different modality data into a common subspace and measures the similarities in the common space. However, with the increasing of data, these traditional methods will suffer from high computing complexity and low search accuracy.…”

Unsupervised Cross-Dataset Person Re-identification by Transfer Learning of Spatial-Temporal Patterns

“…The core of these algorithms is to learn the pedestrian features and the similarity measurements, which are view invariant and robust to the change of cameras. Most of the proposed algorithms [1] [3] [30] [14] [20] [24] conduct supervised learning on the labeled datasets with small size. Directly deploying these trained models to the real-world environment with large-scale camera networks can lead to poor performance, because the target domain may be significantly different from the small training dataset.…”