Selective Multibranch Attention Network With Material Constraint for Baggage Reidentification

Baggage reidentification (re-ID) expects to retrieve the unique baggage across multiple nonoverlapping cameras between the precheck and postcheck areas. This critical challenge task lies in two folds: intraclass dissimilarity by pose variations and interclass similarity due to the identical style. I...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on instrumentation and measurement 2022, Vol.71, p.1-11
Main Authors: Zhang, Hui, Chen, Ruibo, Li, Chen, Chen, Yurong, Wang, Yaonan, Wu, Q. M. Jonathan
Format: Article
Language:English
Subjects:
Adaptation models
Attention mechanism
Baggage
baggage reidentification (re-ID)
Cameras
Computational modeling
Correlation
feature embedding
Feature extraction
material attribute
meta learning
multiple granularities representations
Representations
Task analysis
Visualization
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Baggage reidentification (re-ID) expects to retrieve the unique baggage across multiple nonoverlapping cameras between the precheck and postcheck areas. This critical challenge task lies in two folds: intraclass dissimilarity by pose variations and interclass similarity due to the identical style. In this article, we solve the mentioned problems by a selective multibranch deep network (SMBNet) architecture with the attention mechanism. Specifically, SMBNet incorporates one branch for global feature representations and multiple-part representations corresponding to different local regions. Meanwhile, we plugged the attention block into the part branch to further help the network capture the dependence of the context on local features. Depending on the class response to each part, we propose an adaptive part selection (APS) to automatically select the appropriate part branch and joint with the global one to build a comprehensive baggage descriptor. It effectively blurs unreliable parts and significantly improves representational performance. Finally, we couple classification supervision and material signal in triplet loss to learn a discriminative embedding space that can compact class and material-invariant features and better separate specific representations. Extensive experiments demonstrate the superiority of SMBNet, which achieves 84.7% mean average precision (mAP) and 82.8% Rank-1 on the MVB dataset and significantly outperforms the state-of-the-art methods.
ISSN:0018-9456
1557-9662
DOI:10.1109/TIM.2021.3120799