IAN: The Individual Aggregation Network for Person Search

Abstract

Person search in real-world scenarios is a new challenging computer version task with many meaningful applications. The challenge of this task mainly comes from: (1) unavailable bounding boxes for pedestrians and the model needs to search for the person over the whole gallery images; (2) huge variance of visual appearance of a particular person owing to varying poses, lighting conditions, and occlusions. To address these two critical issues in modern person search applications, we propose a novel Individual Aggregation Network (IAN) that can accurately localize persons by learning to minimize intraperson feature variations. IAN is built upon the state-of-theart object detection framework, i.e., faster R-CNN [1], so that high-quality region proposals for pedestrians can be produced in an online manner. In addition, to relieve the negative effect caused by varying visual appearances of the same individual, IAN introduces a novel center loss that can increase the intraclass compactness of feature representations. The engaged center loss encourages persons with the same identity to have similar feature characteristics. Extensive experimental results on two benchmarks, i.e., CUHK-SYSU and PRW, well demonstrate the superiority of the proposed model. In particular, IAN achieves 77.23% mAP and 80.45% top-1 accuracy on CUHK-SYSU, which outperform the state-of-the-art by 1.7% and 1.85%, respectively.

Learning Context Graph for Person Search

“…To facilitate real-world person re-id, recent methods propose to jointly address the task of detection and reidentification [43,42]. State-of-the-art methods [15,26,41] design online learning object functions to learn large number of identities in the training set. These methods achieve great performance on recent person search datasets.…”

Query-Guided End-To-End Person Search

“…ing training by running averages and allow for employing a soft-max loss in the ID Net training with a limited number of IDs. More recently, [35] extends the OIM with an additional center loss [34], which improves the intra-class feature compactness. To our knowledge, the OIM loss is currently best for optimizing the joint network, adopted by most recent work [23,35], including ours.…”

Simultaneous Multi-View Instance Detection With Learned Geometric Soft-Constraints

“…[33] detects and re-identifies objects in an end-to-end learnable CNN approach with an online instance matching loss. [32] solves re-identification with a socalled center loss that tries to minimize the distance between candidate boxes in the feature space. In contrast to prior work [30,3,15], which does detection, geo-coding and reidentification in a hierarchical procedure, our method does it simultaneously in one pass.…”

“…Methods based on Siamese models [18] alone are not a viable solution to our problem, since they need image crops of the object and can not fully utilize re-identification annotations due to their pairwise labelling training setup. [32] searches for a crop within the detections in a gallery of images, in comparison to our method which aims at matching detections from both full images. The key differences between our work and [33] is that we both ensure object geolocalization and avoid storing features from all identities since that is impractical in a real-world application like the one considered in the paper where objects actually look very similar in appearance.…”