论文学习5——From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization

Abstract

The authors proposed a Training-Free Feature Centralization ReID framework(Pose2ID), and they introduced two components: Identity-Guided Pedestrian Generation and Neighbor Feature Centralization.

Introduction

They utilize identity features of ReID and propose an Identity-Guided Pedestrian Generation paradigm which generate high=quality images of the same person with a high degree of identity consistency across diverse scenarios.

Related works

ReID is a critical task in computer vision that focus on identifying individuals across different camera views.

Feature Enhancement is important for ReID model to differentiate between two people.

Researchers use GANs and diffusion models to generate high-quality pedestrian pose images.

Method

The main purpose of this paper is to centralize features to their identity center.

Identity-Guided Pedestrian Generation

They use a Stable Diffusion model to generate images of the same identity firstly.

Then they proposed the Identity Feature Redistribute(IFR) to utilize identity features from input images.

$H=IFR(f(x))=LN(Linear(f))$

As we can see they also use the ViT structure, including Pose Encoder, VAE Encoder and VAE Decoder.

After extracting the features from different poses, the identity center for IDs will be calculated and to select the image whose feature is the closed to the center. Then 8 representative poses are selected.

After that, they will generate new images for each of these poses.

Neighbor Feature Centralization(NFC)

Finally, they proposed a NFC algorithm to reduce noise in individual features and improve their identity discriminability.