Skip to content

924973292/SD-ReID

BranchesTags

Repository files navigation

SD-ReID: View-aware Stable Diffusion for Aerial-Ground Person Re-Identification

Official PyTorch implementation of the IEEE Transactions on Image Processing (TIP), 2026 paper

Yuhao Wang · Xiang Hu · Lixin Wang · Pingping Zhang · Huchuan Lu

Paper PDF Venue Python PyTorch License

A generative AG-ReID framework that learns view-aware identity representations with Stable Diffusion guidance.

News

  • 2026.05: Codebase cleaned and reproduction guide released for the IEEE TIP 2026 version.
  • 2026: SD-ReID is accepted by IEEE Transactions on Image Processing (TIP).
  • 2025.10: arXiv preview updated to v2: arXiv:2504.09549.

Table of Contents

Overview

Aerial-Ground Person Re-Identification (AG-ReID) aims to retrieve a target person across cameras with sharply different viewpoints. Existing discriminative models often focus on suppressing viewpoint variations, while the view-specific cues themselves can be useful for building stronger identity representations.

SD-ReID introduces a view-aware generative ReID framework. It first trains a ViT-based ReID model to produce person representations and controllable identity/view conditions, then fine-tunes Stable Diffusion to enhance person representations under these conditions. A View-Refined Decoder (VRD) bridges instance-level and global-level features, and the final retrieval uses both person representations and all-view features.

The method is evaluated on five AG-ReID benchmarks: CARGO, AG-ReIDv1, AG-ReIDv2, LAGPeR, and G2APS-ReID. Please refer to the paper for full quantitative results and analysis.

Paper Information

Item Details
Venue IEEE Transactions on Image Processing (TIP), 2026
Preview arXiv:2504.09549, v2 revised on Oct. 30, 2025
Task Aerial-Ground Person Re-Identification
Benchmarks CARGO, AG-ReIDv1, AG-ReIDv2, LAGPeR, G2APS-ReID
Codebase PyTorch, modular ReID training, Diffusers-based Stable Diffusion modules

Method

Component Role
Stage-1 ReID Encoder Learns identity-discriminative and view-aware person representations.
Identity/View Conditions Provide controllable guidance for Stable Diffusion fine-tuning.
Stage-2 SD Module Refines person features through the Stable Diffusion prior.
View-Refined Decoder Aligns instance-level features with global all-view representations.
All-view Retrieval Combines person and all-view features for final ranking.

Installation

conda create -n sdreid python=3.10 -y
conda activate sdreid

# Install the PyTorch build that matches your CUDA driver.
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# Install the remaining dependencies.
pip install -r requirements.txt

Optional Cython acceleration for ranking and ROC evaluation:

cd fastreid/evaluation/rank_cylib
python setup.py build_ext --inplace
cd ../../..

If the Cython extension is unavailable, evaluation falls back to the Python implementation. FAISS is only required when jaccard distance or reranking is enabled.

Pretrained Assets

Place pretrained models under pretrained/ by default:

pretrained/
  jx_vit_base_p16_224-80ecf9dd.pth
  stable-diffusion-v1-5/
    vae/
    unet/
    scheduler/
Asset Expected path Note
ViT-Base ImageNet checkpoint pretrained/jx_vit_base_p16_224-80ecf9dd.pth Used by Stage 1.
Stable Diffusion VAE pretrained/stable-diffusion-v1-5/vae Diffusers format.
Stable Diffusion UNet pretrained/stable-diffusion-v1-5/unet Diffusers format.
Stable Diffusion scheduler pretrained/stable-diffusion-v1-5/scheduler Diffusers format.

For slow Hugging Face access, download with a mirror such as hf-mirror.com or ModelScope and keep the same local folder names.

Datasets

Set a unified dataset root:

export FASTREID_DATASETS=/path/to/datasets

The same path can be passed at runtime with DATASETS.ROOT /path/to/datasets. Dataset files are not included in this repository; please obtain each benchmark according to its license.

Expected dataset layout 
$FASTREID_DATASETS/
  AG_ReID_v2/
    train_all/
    query/
    gallery/
    exp1_aerial_to_cctv.txt
    exp2_aerial_to_wearable.txt
    exp4_cctv_to_aerial.txt
    exp5_wearable_to_aerial.txt
    qut_attribute_v8.mat

  AG_ReID_v1/
    bounding_box_train/
    query_all_c0/
    query_all_c3/
    bounding_box_test_all_c0/
    bounding_box_test_all_c3/
    qut_attribute_v4_88_attributes.mat

  G2APS-ReID/
    bounding_box_train/
    exp3_A2G+.txt
    exp4_G2A+.txt

  LAGPeR/
    bounding_box_train/
    exp3_A2G+.txt
    exp4_G2A+.txt
    exp5_G2A+G.txt

  CARGO/
    train/
    query/
    gallery/

Reproduction

The default seed is fixed to 1111. Stage 2 loads the matching Stage-1 checkpoint by default.

Main AG-ReIDv2 Protocol

python tools/train_net.py --config-file configs/v2/base_stage1.yml --num-gpus 1 \
  DATASETS.ROOT "$FASTREID_DATASETS"

python tools/train_net.py --config-file configs/v2/base_stage2.yml --num-gpus 1 \
  DATASETS.ROOT "$FASTREID_DATASETS" \
  MODEL.WEIGHTS logs/v2/base_stage1_beta/model_best.pth

The same two-stage run can be launched with:

bash run_v2.sh

Experiment Matrix

Benchmark Stage 1 Stage 2 Script
AG-ReIDv2 configs/v2/base_stage1.yml configs/v2/base_stage2.yml bash run_v2.sh
AG-ReIDv1 configs/v1/base_stage1.yml configs/v1/base_stage2.yml bash go/base.sh
G2APS-ReID configs/G2APS_ReID/base_stage1.yml configs/G2APS_ReID/base_stage2.yml bash run_G2APS_ReID.sh
G2APS-ReID 128x64 configs/G2APS_ReID/base_stage1_128.yml configs/G2APS_ReID/base_stage2_128.yml bash run_G2APS_ReID_128.sh
LAGPeR configs/LAGPeR/base_stage1.yml configs/LAGPeR/base_stage2.yml bash run_LAGPeR.sh
CARGO configs/CARGO/base_stage1.yml configs/CARGO/base_stage2_step5.yml bash CARGO.sh

Useful overrides:

# Multi-GPU training
CUDA_VISIBLE_DEVICES=0,1 python tools/train_net.py --config-file configs/v2/base_stage1.yml --num-gpus 2

# Change batch size or output directory
python tools/train_net.py --config-file configs/v2/base_stage1.yml \
  SOLVER.IMS_PER_BATCH 128 OUTPUT_DIR logs/v2/custom_stage1

Evaluation

Evaluate a Stage-2 checkpoint:

python tools/train_net.py --config-file configs/v2/base_stage2.yml --eval-only --num-gpus 1 \
  DATASETS.ROOT "$FASTREID_DATASETS" \
  MODEL.WEIGHTS logs/v2/base_stage2_tstep800/model_best.pth

Evaluate a Stage-1 checkpoint:

python tools/train_net.py --config-file configs/v2/base_stage1.yml --eval-only --num-gpus 1 \
  DATASETS.ROOT "$FASTREID_DATASETS" \
  MODEL.WEIGHTS logs/v2/base_stage1_beta/model_best.pth

Before long runs, a lightweight sanity check is recommended:

python -m compileall fastreid tools
for s in CARGO.sh run_G2APS_ReID.sh run_G2APS_ReID_128.sh run_LAGPeR.sh run_v2.sh go/base.sh; do bash -n "$s"; done

Troubleshooting

Symptom Check
Dataset FileNotFoundError Verify FASTREID_DATASETS or DATASETS.ROOT, and match the folder names above.
Missing ViT checkpoint Put jx_vit_base_p16_224-80ecf9dd.pth under pretrained/, or override MODEL.BACKBONE.PRETRAIN_PATH.
Missing Stable Diffusion files Ensure vae, unet, and scheduler are Diffusers-format folders.
Cython build failure Install cython and a compiler; Python evaluation fallback remains available.
FAISS import failure Install faiss-cpu for jaccard distance/reranking, or keep reranking disabled.
Out of memory Reduce SOLVER.IMS_PER_BATCH, TEST.IMS_PER_BATCH, or TEST.SDMODEL.NUM_INFERENCE_STEPS.

Repository Notes

Generated artifacts and local assets are intentionally excluded from version control: caches, compiled extensions, datasets, checkpoints, pretrained weights, logs, and experiment outputs.

Citation

If SD-ReID is helpful for your research, please cite the paper:

@article{wang2026sdreid,
  title   = {SD-ReID: View-aware Stable Diffusion for Aerial-Ground Person Re-Identification},
  author  = {Wang, Yuhao and Hu, Xiang and Wang, Lixin and Zhang, Pingping and Lu, Huchuan},
  journal = {IEEE Transactions on Image Processing},
  year    = {2026}
}

For the preview version:

@article{wang2025sdreid,
  title   = {SD-ReID: View-aware Stable Diffusion for Aerial-Ground Person Re-Identification},
  author  = {Wang, Yuhao and Hu, Xiang and Wang, Lixin and Zhang, Pingping and Lu, Huchuan},
  journal = {arXiv preprint arXiv:2504.09549},
  year    = {2025}
}

Contact

Questions and discussions are welcome through GitHub Issues. More projects can be found at Yuhao Wang's homepage.

About

【IEEE TIP2026】SD-ReID: View-aware Stable Diffusion for Aerial-Ground Person Re-Identification

Resources

Readme

License

Apache-2.0 license
Activity

Stars

4 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages