Explainable Person Re-Identification with Attribute-guided Metric Distillation
Xiaodong Chen1,2
Xinchen Liu2
Wu Liu2
Xiao-Ping Zhang3
Yongdong Zhang1
Tao Mei2
1University Of Science And Technology Of China
2AI Research of JD.com
3Ryerson University
IEEE International Conference on Computer Vision (ICCV) 2021, Poster Presentation




The motivation of attribute-guided metric distillation. (a) Given a query, the ReID model returns a rank list of gallery images based on pairwise metrics. (b) The learned Interpreter can visualize intuitive attention maps of attributes to tell users what attributes make two persons different, and generate contributions of attributes to reflect the impact of each attribute. (c) Refined results by re-weighted distances from Interpreter. (Best viewed in color.)



Download

Videos at ICCV'2021

Analysis

GitHub Repo


Abstract

Despite the great progress of person re-identification (ReID) with the adoption of Convolutional Neural Networks, current ReID models are opaque and only outputs a scalar distance between two persons. There are few methods providing users semantically understandable explanations for why two persons are the same one or not. In this paper, we propose a post-hoc method, named Attribute-guided Metric Distillation (AMD), to explain existing ReID models. This is the first method to explore attributes to answer: 1) what and where the attributes make two persons different, and 2) how much each attribute contributes to the difference. In AMD, we design a pluggable interpreter network for target models to generate quantitative contributions of attributes and visualize accurate attention maps of the most discriminative attributes. To achieve this goal, we propose a metric distillation loss by which the interpreter learns to decompose the distance of two persons into components of attributes with knowledge distilled from the target model. Moreover, we propose an attribute prior loss to make the interpreter generate attribute-guided attention maps and to eliminate biases caused by the imbalanced distribution of attributes. This loss can guide the interpreter to focus on the exclusive and discriminative attributes rather than the large-area but common attributes of two persons. Comprehensive experiments show that the interpreter can generate effective and intuitive explanations for varied models and generalize well under cross-domain settings. As a by-product, the accuracy of target models can be further improved with our interpreter.


2-Minute presentation video



Architecture
The overall architecture of the attribute-guided metric distillation framework for person ReID. (a) The target ReID model that generates the pairwise distance for an image pair. (b) The interpret network that learns to decompose the pairwise distance into components of attributes and generates attention-guided attention maps for individual attributes. (Best viewed in color.)


Statistics of Attributes

The statistics of attributes in Market-1501 and DukeMTMC-ReID are shown in Figures, which shows that the attributes are very imbalanced
Market-1501
DukeMTMC-ReID


Evaluation and Visualization

(1) Evaluation of interpreters for different backbone models on Market-1501 and DukeMTMC-ReID.


Each target model and the corresponding interpreter are grouped for comparison.

(2) Pairwise examples and explanations for SBS (ResNet-50) on two datasets.


For each pair of images, the upper part visualizes the AAMs of the top-3 attributes, which shows that the AAMs are attended to the discriminative attributes. The lower part shows the overall distance and contributions of the top-3 attributes. These figures show the most contributed attributes discovered by the interpreter.

(3) Under the cross-domain setting?


Evaluation of the interpreters for SBS (ResNet-50) under the cross-domain setting. M to D means the SBS models and interpreters are trained on Market-1501 and tested on DukeMTMC-ReID, and D to M means the reverse setting. The results demonstrate that the information loss of interpreters is very minor under the cross-domain setting.


Download

Log of train and test



TBD
TBD

Updates

[01/10/2021] We include new subsections to track updates and address FAQs.

FAQs

Q0: TBD:
A0: TBD.


Paper

Chen, Liu, Liu, Zhang, Zhang, Mei.
Explainable Person Re-Identification with Attribute-guided Metric Distillation
In ICCV, 2021 (Poster).
(arXiv)
(Additional details/
supplementary materials)


Cite

@inproceedings{chen2021AMD,
title={Explainable Person Re-Identification with Attribute-guided Metric Distillation},
author=Chen, Liu and Liu, Zhang and Zhang, Mei.},
booktitle={IEEE International Conference on Computer Vision (ICCV)},
year={2021}
}


Acknowledgements

This work was supported by the National Key R&D Program of China under Grant No.2020AAA0103800.
This work was done when Xiaodong Chen was an intern at JD AI Research.


Contact

For further questions and suggestions, please contact Xiaodong Chen (cxd1230@mail.ustc.edu.cn).