From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization

Chao Yuan^*  Guiwei Zhang^*  Changxiao Ma  Tianyi Zhang  Guanglin Niu

Beihang University

🔥 A very simple but efficient framework for ReID tasks/models. 🔥

We proposed:

• Training-Free Feature Centralization framework (Pose2ID) that can be directly applied to different ReID tasks and models, even an ImageNet pre-trained model without ReID training;
• Identity-Guided Pedestrian Generation (IPG) paradigm, leveraging identity features to generate high-quality images of the same identity in different poses to achieve feature centralization;
• Neighbor Feature Centralization (NFC) based on sample's neighborhood, discovering hidden positive samples of gallery/query set to achieve feature centralization.

📣 Updates

• [2025.03.06] 🔥 Pretrained weights is available on HuggingFace!
• [2025.03.04] 🔥 Paper is available on Arxiv!
• [2025.03.03] 🔥 Official codes has released!
• [2025.02.27] 🔥🔥🔥 Pose2ID is accepted to CVPR 2025!

⚒️ Quick Start

There are two parts of our project: Identity-Guided Pedestrian Generation (IPG) and Neighbor Feature Centralization (NFC).

IPG using generated pedestrian images to centralize features. Using simple codes could implement:

'''
normal reid feature extraction to get feats
'''
feats_ipg = torch.zeros_like(feats)
# fuse features of generated positive samples with different poses
for i in range(num_poses):
    feats_ipg += reid_model(feats_pose[i]) # Any reid model
eta = 1 # control the impact of generated images (considering the quality)
# centralize features and normalize to original distribution
feats = torch.nn.functional.normalize(feats + eta * feats_ipg, dim=1, p=2) # L2 normalization
'''
compute distance matrix or post-processing like re-ranking
'''

NFC explores each sample's potential positive samples from its neighborhood. It can also implement with few lines:

from NFC import NFC
feats = NFC(feats, k1 = 2, k2 = 2)

📊 Experiments

ID² Metric

We proposed a quantitative metric (ID²) for Identity Density to replce visualization tools like t-SNE, which is random and only focus on few samples.

It can be used in one line:

from ID2 import ID2
density = ID2(feats, pids) # each ID's density
density.mean(0) # global density

where feats is the features extracted by ReID model and pids is the corresponding person IDs.

Improvements on Person ReID tasks

All the experiments are conducted with the offcial codes and pretrained models. We appreciate their official repositories and great works:

• TransReID
• CLIP-ReID
• KPR
• BPBReID
• SAAI
• PMT

Model without ReID training

TransReID loads a ViT pre-trained model on ImageNet for training on the ReID task. This experiment conduct on that pre-trained model which is NOT trained on ReID task.

Ablation Studies

Random Generated Images

🚀 IPG Installation

Download the Codes

git clone https://github.com/yuanc3/Pose2ID
cd Pose2ID/IPG

Python Environment Setup

Create conda environment (Recommended):

conda create -n IPG python=3.9
conda activate IPG

Install packages with pip

pip install -r requirements.txt

Download pretrained weights

1. Download official models from:

   • sd-vae-ft-mse
   • stable-diffusion-v1-5

2. Download our IPG pretrained weights from HuggingFace or Google Drive, and put them in the pretrained directory.

   git lfs install
   git clone https://huggingface.co/yuanc3/Pose2ID pretrained

   The pretrained are organized as follows.

   ./pretrained/
   ├── denoising_unet.pth
   ├── reference_unet.pth
   ├── IFR.pth
   ├── pose_guider.pth
   └── transformer_20.pth
   

Inference

Run the inference.py script. It will generate with poses in the standard_poses for each reference image in ref. The output images will be saved in the output.

python inference.py --ckpt_dir pretrained --pose_dir standard_poses --ref_dir ref --out_dir output

--ckpt_dir: directory of pretrained weights,
--pose_dir: directory of target poses (we provide 8 poses used in our experiment),
--ref_dir: directory of reference images (we provide 10 reference imgs),
--out_dir: directory of output images.

Official generated images on Market1501

Here, we provide our generated images on Gallery and Query of test set on Market1501 with our 8 representative poses.

Getting target poses

We use DWpose to get poses with 18 keypoints.Please follow their official instructions.

📝 Release Plans

Status	Milestone	ETA
🚀	Training codes	TBD
🚀	IPG model trained on more data	TBD
🚀	IPG model with modality transfer ability (RGB2IR)	TBD
🚀	Video-IPG model	TBD