This image shows a research poster from Peking University titled "Distribution-aware Knowledge Prototyping for Lifelong Person Re-identification." The introduction discusses the challenges of lifelong person re-identification (LReID), focusing on issues like catastrophic forgetting and data privacy. The poster proposes Distribution-aware Knowledge Prototyping (DKP) as a solution. Diagrams illustrate concepts such as identity feature centers and statistical feature distributions, with a focus on instance-level diversity for prototyping. A person's hand is visible, likely engaging with the content. The work is affiliated with the Wangxuan Institute of Computer Technology at Peking University.
Text transcribed from the image:
ING
18
大 北京大學
Distribution-aware Knowledge Prototyping for
Kunlun Xu, Xu Zou², Yu
PEKING UNIVERSITY
Introduction
Wangxuan Institute of Computer Technolog
School of Artificial Intelligence and Automation, Huazhong Unive
Method
from
➤ Lifelong person
catastrophic forgetting when learning continually. Exemplar
re-identification (LReID) suffers
and knowledge distillation-based LReID methods encounter
data privacy and limited acquisition capacity respectively.
Prototype-based works set the prototypes as discrete points or
statistical distributions, either discarding the distribution
information or omitting instance-level diversity which are
crucial fine-grained clues for LReID.
We propose Distribution-aware Knowledge Prototyping (DKP)
where instance-level diversity of each sample is modeled to
transfer comprehensive fine-grained knowledge for prototyping.
Identity Feature
Identity Center
-Step t-1
-Step t-
Dt-1
Pt-1
Dâ‚‚
(a) Overall Model
X1 X2
Χη
Backbone
Instance center Prototype of step t-1
H
Pooling
H
P-10
PKT
IDM
H
Lproto-di
Lce-d
Liri-di
DPG
P
Conv Linear
Model Mt Instance variance Prototype of step t
Distribution Boundary
(c) Prototype-based Knowledge Transfer (PKT)
91 1
(b) Inst
Insta
New
Identity
Step t-1
(a) Identity feature center A as prototype
New
Identity
Step t-1
(b) Statistical feature distribution (Aas prototype
Step t-1
New
Identity
Step t
HN=-1
Prototype Pt-1
Sampler Prototype sa
➤ Overview: Our model is buil
distribution center and va
IDM and DPG are design
➤IDM: Model the distr
proposed distribut
Step t
➤ PKT: Guide the no
Step t
(c) Instance-level distribution generated prototype (Ours)
A
the aid of historical distribu.
Cp = P(Fc F/2₁)
DPG: Transform the instance-level
Gaussian distribution which is regis
nk
1
14=-
Ci
nk
i=1