This image displays a detailed research poster titled "Learning Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification." The poster, presented at CVPR (Computer Vision and Pattern Recognition) Conference in Seattle, WA, from June 17-21, 2024, showcases the work by Zhenyu Cui, Jiahuan Zhou, Xun Wang, Manyu Zhu, and Yixin Peng from ByteDance Inc and Wangxuan Institute of Computer Technology, Peking University.

The poster elaborates on innovative techniques for Lifelong Person Re-identification (L-ReID), emphasizing the Continual Compatible Representation (C²R) scheme. The methodology addresses the re-indexing challenge by employing strategies such as Continual Compatible Transfer (CCT), Balanced Compatible Distillation (BCD), and Balanced Anti-forgetting Distillation (BAD).

The 'Methods' section outlines these approaches, and detailed diagrams demonstrate various stages and workflows involved. 

The 'Experiments' section contains comparative analysis tables reflecting the superior performance of the proposed C²R on several benchmark L-ReID datasets. Graphs further illustrate the stable SOTA (state-of-the-art) performance across different training stages.

The poster summarizes its findings in the 'Conclusion' section, proposing C²R for practical L-ReID applications and highlighting the balance achieved between anti-forgetting and compatibility in new feature accumulation.

Additional QR codes link to a lab homepage, GitHub page, and a WeChat page for more in-depth information and related works.
Text transcribed from the image:
Byte Dance
Introduction:
Learning Continual Compatible Representation for Re-indexing Free Lifelong Person Re-identification
Zhenyu Cui', Jiahuan Zhou, Xun Wang2, Manyu Zhu², Yuxin Peng1*
"Wangxuan Institute of Computer Technology, Peking University 2ByteDance Inc
Lifelong Person Re-identification (L-ReID): Aims to learn from
sequentially collected data to match a person across different
scenes. Once the model is updated, historical images in the gallery
are re-calculated to obtain new features for testing (Re-indexing).
Feature Replacing
RelD w/Re-indexing
RelD w/o Re-indexing
DO
Privacy
O Gallery Features
Query Fest
• Limitations: Re-indexing is infeasible when raw images in the
gallery are unavailable due to data privacy concerns, resulting in
incompatible retrieval between the query and the gallery features
calculated by different models and limitation on L-ReID performance.
Goal: Achieve effective L-ReID
Method:
• Overview: A Continual Compatible Representation (C2R) scheme is
proposed for RFL-ReID to continuously update old features in the
gallery to make it compatible with new query features.
> Continual Compatible Transfer (CCT): Update old gallery
features continuously and transfer old feature to the new feature
space, adaptively capturing the knowledge from different domains.
➤ Balanced Compatible Distillation (BCD): Achieve the
compatibility between the transferred features and the new ones,
preserving the relationship between the old and the transferred
features in a unified feature space.
➤ Balanced Anti-forgetting Distillation (BAD): Eliminate the
accumulated forgetting of old knowledge during the continuous
transfer, balancing the old and the new discriminative information.
( Training at Stepe
Continual Compatible Transfer (CCT)
Balanced Compatible Distillation (BCD)
Balanced Anti-forgetting Distillation (BAD)
Experiments:
The CVPR
M SEATTLE, WA JUNE 17-21, 2024
•
Task
Method
mAP
MAP R@I
R@I
Market-1501 CUHK-SYSU DukeMTMC-ReID MSMT17-V2
mAP ROI MAP ROI
The experiment on the benchmark L-ReID datasets verify the
superiority of our proposed C²R.
Joint Train
68.1
85.2
81.4 83.8
60.4
75.7
24.6 48.9
Average
CUHK03
MAP R@1 mAP R@I
42.7 43.6 55.4 67.5
SPD [28]
35.6
61.2
61.7
64.0
27.5
47.1
5.2
15.5
44.3
42.2
34.4 46.4
LwF [11]
563
56.3
77.1
72.9
75.1
29.6
46.5
6.0 16.6
36.1
37.5
40.2
50.6
CRE
CRL [40]
58.0
78.2
72.5
75.1
28.3
45.2
6.0
15.8
37.4
39.8
40.5
50.8
L-ReID
AKATIO
AKA [18]
58.1
77.4
72.5
74.8
28.7
45.2
6.1
38.7
16.2
40.4
40.8 50.8
MEGE [20]
39.0 01.0
61.6
73.3
76.6
16.9
30.3
4.6
13.4
36.4
37.1
34.0
43.8
PatchKD [27] 68.5
85.7
75.6
78.6
33.8
50.4
17.0
6.5
34.1
36.8
43.7
53.7
RFL-ReID
Ours
58.0 40.0
39.1 58.0
LWF [11]
52.2
36.1
AKA [18]
49.0
55.6
38.8
CVS* [30]
78.4 57.8
PatchKD [27] 61.4
Ours 62.7 79.7 64.4
86.8 76.7
69.0
79.5
33.2
48.6
35.6
17.4
6.6
36.2
44.2
53.7
40.7
7.8
15.3
2.6
7.1
23.3 23.9 22.6 29.0
37.6
38.6
7.6
13.8
3.1
8.3
26.5
26.5 22.4
27.7
49.7 19.3
30.0
4.6
20.8
59.0
26.7
66.3
34.4
5.1
41.7
24.7 24.7 27.3
11.5
12.8 36.0 37.6 36.2
6.8 15.7 37.2 37.6 39.5
34.3
44.4
48.2
Our C2R achieves stable SOTA performance at each training stages.
Contal
Compatible
without re-indexing raw images
in the gallery by continuously
updating old features in the
Regrestative
gallery to make it compatible
with new query features.
• Contribution:
日日"
➤Focus on a practical but challenge problem on L-ReID task, called
Re-indexing Free Lifelong Person Re-identification (RFL-ReID).
Propose a Continual Compatible Representation (C2R) scheme to
prevent re-indexing to raw images in the old gallery in L-ReID.
A Continual Compatible Transfer (CCT) and two balanced
distillation modules are proposed to achieve L-ReID without re-
indexing raw images in the gallery.
Transfer at Stage
Pipeline:
(c) Testing at Stage
ID: 112V
Query
Training Stage: CCT, BCD, and BAD are jointly optimized.
Transferring Stage: CCT network is employed to update the old
feature set after each training stage in L-ReID.
➤ Testing Stage: Query features calculated by the new model is
used to directly match the updated gallery features to achieve re-
indexing free lifelong person re-identification.
➤Three stages are executed sequentially when facing new data.
Conclusion
•
.
•
We propose C2R for the practical & challenge task, RFL-ReID,
tackling the data privacy issue in lifelong person re-identification.
Two balanced distillation module are designed to balance the anti-
forgetting of old knowledge with the compatibility to the new model.
Extensive experiments on several benchmark L-ReID datasets
demonstrate the effectiveness of our method.
More
Information
Go!
Related
Works
Go!
[Lab homepage]
[GitHub page]
MIPL
[WeChat page]
[1] Zhenyu Cui, et al., "Continual Vision-Language Retrieval via Dynamic Knowledge
Rectification", 8th AAAI Conference on Artificial Intelligence (AAAI), 2024.
[2] Zhenyu Cui, et al., "DMA: Dual Modality-Aware Alignment for Visible-Infrared Person
Re-Identification", IEEE Transactions on Information Forensics and Security (TIFS), 2024.