Pyramidal Person Re-IDentification via Multi-Loss Dynamic Training

By Feng Zheng, Cheng Deng, Xing Sun, Xinyang Jiang, Xiaowei Guo, Zongqiao Yu, Feiyue Huang, Rongrong Ji. Computer Vision and Pattern Recognition, 2018

392 citations recorded by the source corpus.

Most existing Re-IDentification (Re-ID) methods are highly dependent on precise bounding boxes that enable images to be aligned with each other. However, due to the challenging practical scenarios, current detection models often produce inaccurate bounding boxes, which inevitably degenerate the performance of existing Re-ID algorithms. In this paper, we propose a novel coarse-to-fine pyramid model to relax the need of bounding boxes, which not only incorporates local and global information, but also integrates the gradual cues between them. The pyramid model is able to match at different scales and then search for the correct image of the same identity, even when the image pairs are not aligned. In addition, in order to learn discriminative identity representation, we explore a dynamic training scheme to seamlessly unify two losses and extract appropriate shared information between them. Experimental results clearly demonstrate that the proposed method achieves the state-of-the-art results on three datasets. Especially, our approach exceeds the current best method by 9.5% on the most challenging CUHK03 dataset.

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/545e3f4e684f82eaf211bde9558d00a0dad946d2

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

Dynamic Dual-Attentive Aggregation Learning for Visible-Infrared Person Re-Identification (2020, methodology): “Person re-identification (Re-ID) techniques [59,68] have achieved human-level performance with part-level deep feature learning [4,40,67].”
A Strong Baseline and Batch Normalization Neck for Deep Person Re-Identification (2019, background): “Pyramid [51] achieves surprising performance on two datasets, but it concatenates 21 local features of different scales.”
ProxyNCA++: Revisiting and Revitalizing Proxy Neighborhood Component Analysis (2020, background): “It is often applied to images, and has found numerous applications such as visual products retrieval [16,23,1], person re-identification [37,30], face recognition [22], few-shot learning [28,14], and clustering [11].”
Multi-Task Learning with Deep Neural Networks: A Survey (2020, methodology): “Following (Kendall et al., 2017), several methods of weighting multi-task loss functions were introduced which weigh a task’s loss by the learning speed on that task (Chen et al., 2017; Liu et al., 2019; Zheng et al., 2018; Liu et al., 2019a), with slightly varying approaches.”
Identity-Guided Human Semantic Parsing for Person Re-Identification (2020, methodology): “Some researchers directly partition the person image into rigid horizontal stripes to learn local features [38, 43, 52, 55].”
Deep Learning for Person Re-Identification: A Survey and Outlook (2020, methodology): “In addition, a dynamic multi-loss training strategy is designed in [162] to adaptively integrated two loss functions.”
Batch Coherence-Driven Network for Part-Aware Person Re-Identification (2020, background): “For example, BCD-Net scores 86.2% and 81.6% on Rank-1 accuracy and mAP respectively on CUHK03-Labeled; this represents an improvement over Pyramid [20] of 7.3% in terms of Rank-1 accuracy and 4.7% in terms of mAP.”
Channel Augmented Joint Learning for Visible-Infrared Recognition (2021, background): “Visible-Infrared Person Re-IDentification (VI-ReID) is a cross-modality person recognition problem, which aims at matching daytime visible and night-time infrared images [2, 48, 16, 57].”
Lifelong Person Re-Identification via Adaptive Knowledge Accumulation (2021, background): “Note that SFT-T and CRL-T (with additional triplet loss) is not beneficial for the first three training steps, because when the number of training identities is large enough, triplet loss contributes less on performance and even leads to conflict with cross-entropy loss [51].”