Structured Anchor-Guided Aggregation for ReID

Official training/evaluation code for the paper.

Paper: From Global to Local: Rethinking CLIP Feature Aggregation for Person Re-Identification

What is included

This open-source package includes the standalone training/evaluation pipeline:

• train.py is the main entrypoint

Environment setup

Tested with Python 3.8+, PyTorch 2.4, CUDA 12.

1. Create and activate an environment

Option A (Conda, recommended):

conda create -n saga-reid python=3.10 -y
conda activate saga-reid

Option B (venv):

python -m venv .venv
# Linux/macOS
source .venv/bin/activate
# Windows (PowerShell)
.\.venv\Scripts\Activate.ps1

2. Install PyTorch

Install the PyTorch build that matches your system and CUDA version.

CUDA 12.1 example:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

CPU-only example:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

3. Install project dependencies

pip install -r requirements.txt

4. Verify installation

python -c "import torch; print('torch:', torch.__version__); print('cuda_available:', torch.cuda.is_available())"
python -c "import timm, ftfy, regex, PIL, scipy, sklearn, numpy; print('dependencies_ok')"

Supported datasets

• market1501
• dukemtmc
• occ_duke
• occ_market
• occ_reid
• mmmp

Training

python train.py --config configs/market1501.json

Switch to old v1-style features (config only)

Set this in your config file:

"use_v1_features": true

Resume

# resume from Stage-2 checkpoint
python train.py --config configs/market1501.json --start_stage 3

# start Stage 3b directly
python train.py --config configs/market1501.json --start_stage 3 --start_s3b

Stage 4 training is disabled by default ("enable_stage4": false).

Evaluation

# fusion (default, using Stage-3 checkpoint)
python train.py --config configs/market1501.json --eval output/structured_anchor_guided/model_stage3_best.pth

Repository layout

train.py
configs/
  market1501.json
  occ_market1501.json
  occ_dukemtmcreid.json
datasets/
loss/
model/
utils/
visual_modules.py