A detailed academic poster presented at a conference, created by Peking University. It explores "Distribution-aware Knowledge Prototyping for Non-exemplar Lifelong Person Re-identification". The poster includes intricate diagrams, charts, and text discussing methods for Lifelong Person Re-Identification (LReID), highlighting strategies to overcome challenges such as catastrophic forgetting, limited acquisition capacity, and instance-level diversity. An individual points to a section of the poster, perhaps elaborating on experimental results or specific aspects of their proposed methodology. The research involves collaboration with the Wangxuan Institute of Computer Technology, School of Artificial Intelligence and Automation at Huazhong University of Science and Technology, located in China.
Text transcribed from the image:
9
北京大學
PEKING UNIVERSITY
Distribution-aware Knowledge Prototyping for Non-exemplar Lifelong Person Re-identification
Kunlun Xu, Xu Zou², Yuxin Peng, Jiahuan Zhou
School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, 430074, China?
Wangxuan Institute of Computer Technology, Peking University, Beijing 100871, China
ction
suffers from
g person re-identification (LReID)
Ophic forgetting when learning continually. Exemplar
owledge distillation-based LReID methods encounter
vacy and limited acquisition capacity respectively.
e-based works set the prototypes as discrete points or
al distributions, either discarding the distribution
ion or omitting instance-level diversity which are
ne-grained clues for LReID.
pose Distribution-aware Knowledge Prototyping (DKP)
stance-level diversity of each sample is modeled to
comprehensive fine-grained knowledge for prototyping.
Identity Feature
Identity Center
Distribution Boundary
New
Identity
Step t-1
Step t
ty feature center A
as prototype
A
Method
SEATTLE
Step t-1.
Dt-1
-Step t.
Experiments
Pt-1
Dt
Pt
-Step t+1
D+1
Peti
The experiments on the LReID benchmark
Method
Marke
MTMC MSMT17
mAP Re
(a) Overall Model
Instance center Prototype of step t-1
(b) Instance-level Distribution Modeling (IDM)
Instance center
Joint-Train
75.3 91
316 57
L&F (21
CIL SPD [31]
56.3 77
Pt-10
0
Lee-d
PKT
Pooling
IDM
He-d
Lproto-di
Liri-di
Xn
X1 X2
Backbone
H
DPG
.....
Liri-d
CRL (42)
N(CV)
Conv
Linear
Pt
Model Mt
Instance variance Prototype of step t
Instance variance
Sampler
m candidates per age
Feature candidate
(d) Distribution-oriented Prototype Generation (DPG)
35.6 6
PRAKA 12 374 61
PRD [2] 73 180
580 782 725 75.1
AKA [24] 512 720 475 451 187 33.1
LReID AKA [24] 581 714 72.5 748 287 452 61
PatchKD [25] 68.5 85.7 75.6 78.6 33.8 504 65
MEGE [26] 390 616 73.3 76.5 16.9 30.3 49
DKP(Ours) 603 805 83.6 85.4 51.6 68.4 10
Our DKP achieves better balance between learning
> Performance tendency on seen and unse
(c) Prototype-based Knowledge Transfer (PKT)
91 1
2
92 proto-d
Instances Identity i
UN₁-1
Prototype Pt-1
Sampler Prototype sample
Instance center Instances of identity k
Eq. (4) and Eq. (5) Prototype P
Overview: Our model is built upon a dual branch convolution network where the instance-specific
distribution center and variance are predicted, respectively. Furthermore, three modules PKT,
IDM and DPG are designed to accomplish the lifelong learning procedure.
IDM: Model the distribution for each input instance based on the sampling strategy and the
proposed distribution-aware losses ce-d and Ltri-d
Lce-d
1
m
==
m+1
j=1
Taining stop
+a+D
+4
<+-LNE
(a) Anti-
Ltri-d = log(1+ exp(c-fp-f
Ou
th
I learning via enhancing the discriminant of new identity features with
I on the prototype-aware knowledge transfer loss proto-d
=LKL (P(CC/2)p(FF/2))
a multiva