IEEE 2017 Conference on Computer Vision and Pattern Recognition Gallery set from CAM B CAM A Ranking list from CAM B CAM A Pose Background Occlusion Light Viewpoint Scale Definition of person re-identification Contributions Challenges of person re-identification Our method Experimental results Personal information Conv1_1 Conv1_2 Conv1_3 Concat Conv2_1 Conv2_2 Conv2_3 Concat FC_body MSCAN MSCAN MSCAN MSCAN FC_part2 Feature fusion MSCAN Conv0 Conv3_1 Conv3_2 Conv3_3 Concat Conv4_1 Conv4_2 Conv4_3 Concat FC_part3 FC_part1 FC_part Identification FC_loc θ_part1 Grid generator MSCAN FC_loc θ_part2 Grid generator MSCAN FC_loc θ_part3 Grid generator Spatial Transformer Networks Spatial Transformer Networks Spatial Transformer Networks Concat Part #1 Part #2 Part #3 Prior constraints Object function Learning Deep Context-aware Features over Body and Latent Parts for Person Re-identification Dangwei Li, Xiaotang Chen, Zhang Zhang, Kaiqi Huang {dangwei.li, xtchen, zzhang, kqhuang}@nlpr.ia.ac.cn 1) An efficient multi-scale context-aware network for fine-grained pedestrian feature learning. 2) A latent part localization network to atomically localize pedestrian parts without explicitly human parts supervision. 3) An efficient pedestrian representation by fusing global full body and local body parts for pedestrian retrieval. 1) Prior constraints of body part’s center 2) Prior constraints on value range of scale 3) Prior constraints of cropped body parts 4) Prior constraints on localization 5) Identification loss 6) Final object function Visualization of original image, rigid parts, and learned latent body parts Market-1501 CUHK03-labeled CUHK03-detected MARS (Single query) Cross-dataset ReID on VIPeR MARS (multiple query) Visualization of query image and its corresponding retrieval results (body, parts, fusion) Name: Dangwei Li (党伟李) Tel： +86 15600616646 Homepage: http://dangweili.github.io

IEEE 2017 Conference on Learning Deep Context-aware Features over Body … · 2018. 1. 23. · t FC_body MSCAN MSCAN MSCAN MSCAN FC_part2 e fusion MSCAN Conv0 Conv3_1 Conv3_2 Conv3_3

Download PDF Report

Upload
others
View
0
Download
0

Embed Size (px)

Citation preview

IEEE 2017 Conference on

Computer Vision and Pattern

Recognition

Gallery set from CAM B

CAM A Ranking list from CAM B

CAM A

PoseBackground Occlusion Light ViewpointScale

Definition of person re-identification

Contributions

Challenges of person re-identification

Our method Experimental results

Personal information

Conv1_1

Conv1_2

Conv1_3

ncat

Conv2_1

Conv2_2

Conv2_3

ncat

FC_body

MSCANMSCAN

MSCAN

FC_part2

Feature

fusio

MSCAN

Conv0

Conv3_1

Conv3_2

Conv3_3

ncat

Conv4_1

Conv4_2

Conv4_3

ncat

FC_part3

FC_part1

FC_part

Iden

tificationFC_loc θ_part1

Grid generator

MSCAN FC_loc θ_part2Grid

generator

MSCAN FC_loc θ_part3Grid

generator

Spatial Transformer Networks

ncat

Part #1

Part #2

Part #3

Prio

r con

straints

Object function

Learning Deep Context-aware Features over Body and Latent Parts for Person Re-identificationDangwei Li, Xiaotang Chen, Zhang Zhang, Kaiqi Huang

{dangwei.li, xtchen, zzhang, kqhuang}@nlpr.ia.ac.cn

1) An efficient multi-scale context-aware network for fine-grained pedestrian feature learning.2) A latent part localization network to atomically localize pedestrian parts without explicitly human parts supervision.3) An efficient pedestrian representation by fusing global full body and local body parts for pedestrian retrieval.

1) Prior constraints of body part’s center

2) Prior constraints on value range of scale

3) Prior constraints of cropped body parts

4) Prior constraints on localization

5) Identification loss

6) Final object function

Visualization of original image, rigid parts, and learned latent body parts

Market-1501 CUHK03-labeled CUHK03-detected

MARS (Single query) Cross-dataset ReID on VIPeRMARS (multiple query)

Visualization of query image and its corresponding retrieval results (body, parts, fusion)

Name: Dangwei Li (党伟李)Tel： +86 15600616646Homepage: http://dangweili.github.io