Glad

Abstract

The huge variance of human pose and the misalignment of detected human images signi cantly increase the di culty of person Re-Identi cation (Re-ID). Moreover, e cient Re-ID systems are required to cope with the massive visual data being produced by video surveillance systems. Targeting to solve these problems, this work proposes a Global-Local-Alignment Descriptor (GLAD) and an e cient indexing and retrieval framework, respectively. GLAD explicitly leverages the local and global cues in human body to generate a discriminative and robust representation. It consists of part extraction and descriptor learning modules, where several part regions are rst detected and then deep neural networks are designed for representation learning on both the local and global regions. A hierarchical indexing and retrieval framework is designed to eliminate the huge redundancy in the gallery set, and accelerate the online Re-ID procedure. Extensive experimental results show GLAD achieves competitive accuracy compared to the state-of-theart methods. Our retrieval framework signi cantly accelerates the online Re-ID procedure without loss of accuracy. Therefore, this work has potential to work better on person Re-ID tasks in real scenarios.

FSRM-STS: Cross-dataset pedestrian retrieval based on a four-stage retrieval model with Selection–Translation–Selection

“…The methodologies are also relevant to many other image retrieval tasks, and similar researches have been verified in many scenarios, such as buildings, vehicles, devices, and face identification because the methodologies use representation learning to automatically extract image features based on different retrieval tasks. In future, more technologies will be attempted to incorporate, such as local feature learning [4,28,31,33,34] and attribute learning [13,40], to further enhance the capability of the proposed model in Pedestrian Retrieval, and to extend the generalizability to other scenarios.…”

“…Since Krizhevsky [29] won the ILSVRC-2012 competition, the CNN-based deep learning model has become popular. In recent years, some representative deep pedestrian retrieval models have included image slice [30], human body recognition [31,32,33,34], a combination of long-term and shortterm memory networks [35,36,63], and GAN to generate more samples [37,38,39] etc. Based on the existing researches, plenty of innovations have been proposed in recent years to continuously improve the performance on benchmark datasets.…”

A Mask-Pooling Model With Local-Level Triplet Loss for Person Re-Identification

“…Among deeply learned features, part features draw increasing attention and have been proved to be more discriminative [18], [30], [32], [33]. For instance, [14], [32], [34], [35] employ pose information to help part feature learning, and [13], [15]- [20] partition pedestrians into several parts (horizontal stripes, rectangle blocks, etc.) to extract part features.…”

“…The most challenging problem of the above pose-driven methods and partition methods lies in body part misalignment [36]. Body joints [12], pose boxes [34], keypoints [13] and semantic features of different body regions [12], [14] are effective means to address this problem. Inspired by these methods, we propose a novel model that divides pedestrian into several parts.…”

“…2 (a)-(c). One of our intuitions is that, FIGURE 2: (a) -(c) are partition methods in part feature learning, which still contain large amounts of background information: (a) horizontally segmentation [16], [20], (b) key points and part regions [13], (c) body regions and joints [12]. (d) illustrates the motivation of our method: to remove the extra background using mask information.…”