Harmonious Attention Network for Person Re-identification

By Xiatian Zhu, Shaogang Gong. 2018

1,451 citations recorded by the source corpus.

Existing person re-identification (re-id) methods either assume the availability of well-aligned person bounding box images as model input or rely on constrained attention selection mechanisms to calibrate misaligned images. They are therefore sub-optimal for re-id matching in arbitrarily aligned person images potentially with large human pose variations and unconstrained auto-detection errors. In this work, we show the advantages of jointly learning attention selection and feature representation in a Convolutional Neural Network (CNN) by maximising the complementary information of different levels of visual attention subject to re-id discriminative learning constraints. Specifically, we formulate a novel Harmonious Attention CNN (HA-CNN) model for joint learning of soft pixel attention and hard regional attention along with simultaneous optimisation of feature representations, dedicated to optimise person re-id in uncontrolled (misaligned) images. Extensive comparative evaluations validate the superiority of this new HA-CNN model for person re-id over a wide variety of state-of-the-art methods on three large-scale benchmarks including CUHK03, Market-1501, and DukeMTMC-ReID.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for deep neural networks (2020, methodology): “While sophisticated learning based partitioning methods have been explored [36, 37, 38, 39], methods such as [40] have achieved outstanding performance with trivial deterministic splitting.”
Face Anti-Spoofing: Model Matters, so Does Data (2019, methodology): “Unlike traditional strategies like random choice [2] or simple fusion of adding convolutional layer [24], we carefully design an attention mechanism, RAM, with fewer parameters and more reasonable initialization settings for locating the discriminative and significant sub-regions.”
Learning to Reduce Dual-Level Discrepancy for Infrared-Visible Person Re-Identification (2019, background): “Person re-identification (re-ID) has recently received increasing attention in the computer vision community [18, 17, 6, 7, 20, 28] because of its great importance in video surveillance.”
ABD-Net: Attentive but Diverse Person Re-Identification (2019, methodology): “Unless otherwise specified, we use ResNet-50 as the default backbone network due to its popularity in Re-ID [45, 46, 47, 48, 49, 11, 50, 51].”
Memory-Based Neighbourhood Embedding for Visual Recognition (2019, neutral): “0 HA-CNN [14] CVPR’18 41.”
Deep Learning for Person Re-Identification: A Survey and Outlook (2020, methodology): “1) For attention within the person image: Harmonious Attention CNN (HA-CNN) model [79] jointly learns the soft pixel attention and hard regional attention to enhance the robustness of feature representation against misalignment.”
Continual egocentric object recognition (2019, methodology): “We performed these tests on a closed-world re-identiﬁcation setting, similar to the one used in person re-identiﬁcation [13, 35].”
Transferable Joint Attribute-Identity Deep Learning for Unsupervised Person Re-identification (2018, methodology): “Person Re-ID Most existing re-id models are based on supervised learning for every camera pair on a separate set of labelled training data [23, 58, 63, 55, 60, 52, 56, 8, 29, 50, 57, 48, 6, 30, 7, 31, 5].”
Self-Referenced Deep Learning (2018, background): “Deep neural networks have been shown to be effective for solving many computer vision tasks [22, 39, 42, 12, 24, 31].”
Identity-Guided Human Semantic Parsing for Person Re-Identification (2020, methodology): “Attention mechanism constructs alignment by suppressing background noise and enhancing the discriminative regions [3, 23, 31, 41, 50, 58].”
Dynamic Dual-Attentive Aggregation Learning for Visible-Infrared Person Re-Identification (2020, methodology): “Some work [22,26,41] has also investigated using multi-scale or different convolutional channels to capture the pixel-level/small-region-level attentions [35,36].”
Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-Identification (2019, background): “The art supervised methods have made great achievement in person re-ID [14, 26, 39, 25], relying on rich-labeled data and the success of deep networks [18, 12, 8].”
Adaptive Transfer Network for Cross-Domain Person Re-Identification (2019, background): “Deep learning based methods [47, 14, 26, 15, 40] for feature extraction have shown substantial advantage over traditional hard-crafted features on most of person reidentification datasets.”
Pyramidal Person Re-IDentification via Multi-Loss Dynamic Training (2018, methodology): “…[6] 65.5 83.7 - - PAN [37] 63.4 82.8 - - SVDNet [23] 62.1 Pyramid-ours 79.0 89.0 MGN [26] 78.4 88.7 SVDNet [23] 56.8 76.7 AOS [8] 62.1 79.2 HA-CNN [14] 63.8 80.5 GSRW [20] 66.4 80.7 DuATM [21] 64.6 81.8 PCB+RPP [24] 69.2 83.3 PSE+ECN [19] 75.7 84.5 DNN-CRF [3] 69.5 84.9 GP-reid [28] 72.8 85.2…”
Pose-Guided Feature Alignment for Occluded Person Re-Identification (2019, background): “Although there exist occluded person images, the holistic images in DukeMTMC-reID are much more than the occluded ones, so that this dataset can be treated as a holistic re-id dataset in previous works [46, 17, 32].”
Context-aware Feature Generation For Zero-shot Semantic Segmentation (2020, background): “Although our contextual module looks similar to channel attention [14, 23] or full attention [41], our motivation and technical details are intrinsically different from them.”