person re-identification - github pages

1
VP-ReID: Vehicle and Person Re-Identification System Longhui Wei, Xiaobin Liu, Jianing Li, Shiliang Zhang School of Electronic Engineering and Computer Science, Peking University, Beijing 100871, China First offline clusters similar images into same group Images in returned groups are retrieved with original GLAD to generate an image rank list. Offline Grouping and Online Retrieval Re-ID Efficiency on Market1501 Vehicle Re-Identification: GLAD Descriptor GLAD explicitly leverages local and global cues to generate a discriminative and robust representation Learn complementary features on both coarse-grained local parts and global regions The matching of local representation can effectively handle the misalignment and pose change issues Person Re-Identification: Abstract Person Re-Identification (ReID) and vehicle ReID are the key technology in smart surveillance system. We develop a robust and efficient ReID system, named VP-ReID, to demonstrate our recent research progresses on those two tasks. This system is build based on our recent works including discriminative feature design and efficient off-line indexing. Constructed upon those algorithms, VP-ReID can identify vehicle and person efficiently and accurately from large gallery set. Comparison with Related Works On Market1501 On DukeMTMC-ReID On CUHK03 Region-Aware deep Model (RAM) We propose a Region-Aware deep Model (RAM) to jointly learn global and regional features. Attribute cues are additionally used to jointly train RAM. Learned features are more discriminative to detailed local cues, and contains attribute cues. Vehicle ID Attributes (512) classifier Attribute branch Pool (3×6×512) Pool (3×6×512) Pool (4×6×512) FC (1024) FC (1024) FC (1024) 1024 1024 b 1024classifier classifier classifier Region branch Pool (6×6×512) FC (4096) 1024/409 6classifier Conv branch Pool (6×6×512) FC (4096) 1024/409 6BN classifier BN branch Shared convolutional layers Dataset & Code Experimental Results Query Query Query Query Retrieval Result

Upload: others

Post on 10-Dec-2021

2 views

Category:

Documents


0 download

TRANSCRIPT

VP-ReID: Vehicle and Person Re-Identification SystemLonghui Wei, Xiaobin Liu, Jianing Li, Shiliang Zhang

School of Electronic Engineering and Computer Science, Peking University, Beijing 100871, China

◆ First offline clusters similar images into same group

◆ Images in returned groups are retrieved with original GLAD to

generate an image rank list.

Offline Grouping and Online Retrieval

Re-ID Efficiency on Market1501

Vehicle Re-Identification:

GLAD Descriptor

◆ GLAD explicitly leverages local and global cues to generate a

discriminative and robust representation

◆ Learn complementary features on both coarse-grained local

parts and global regions

◆ The matching of local representation can effectively handle the

misalignment and pose change issues

Person Re-Identification:

AbstractPerson Re-Identification (ReID) and vehicle ReID are the key

technology in smart surveillance system. We develop a robust and

efficient ReID system, named VP-ReID, to demonstrate our recent

research progresses on those two tasks. This system is build based

on our recent works including discriminative feature design and

efficient off-line indexing. Constructed upon those algorithms,

VP-ReID can identify vehicle and person efficiently and

accurately from large gallery set.

Comparison with Related Works

On Market1501 On DukeMTMC-ReID

On CUHK03

Region-Aware deep Model (RAM)

◆ We propose a Region-Aware deep Model (RAM) to jointly

learn global and regional features.

◆ Attribute cues are additionally used to jointly train RAM.

◆ Learned features are more discriminative to detailed local cues,

and contains attribute cues.

Vehicle ID

Attributes

𝑓𝑎(512)

classifier

Attribute branch

Pool(3×6×512)

Pool(3×6×512)

Pool(4×6×512)

FC(1024)

FC(1024)

FC(1024)

𝑓𝑟𝑡(1024)

𝑓𝑟𝑚(1024)

𝑓𝑟b(1024)

classifier

classifier

classifier

Region branch𝑅𝑡

𝑅𝑚

𝑅𝑏

Pool(6×6×512)

FC(4096)

𝑓𝑐(1024/409

6)classifier

Conv branch

Pool(6×6×512)

FC(4096)

𝑓𝑏(1024/409

6)BN classifier

BN branch

Shared convolutional layers

𝑀

Dataset & Code

Experimental Results

Query Query

Query Query

Retrieval Result