layumi/Person_reID_baseline_pytorch

What is this repo?

Person ReID Baseline (PyTorch) is an open‑source code base for object‑re‑identification – the task of matching images of the same person (or other objects) across different camera views. It provides a compact, ready‑to‑run implementation of many tricks that have become standard in the research literature since 2017, all built on PyTorch.

---

Key points (from the README)

Aspect	Details
Goal	Offer a strong (state‑of‑the‑art performance), small (fits in ~2 GB GPU memory with bf16/fp16), and friendly (easy to turn on/off tricks) baseline for person re‑identification.
Performance	Using only a softmax loss the baseline reaches Rank@1 ≈ 88 %, mAP ≈ 71 % on the Market‑1501 benchmark.
Supported models	ResNet, ResNet‑IBN, DenseNet, HRNet, EfficientNet, Swin Transformer, Swin‑V2, ConvNeXt, DinoV3, and more.
Losses	Cross‑entropy, Circle, Triplet, Contrastive, Sphere, Lifted, ArcFace, CosFace, Instance, and others.
Training tricks	bf16/fp16, random erasing, linear warm‑up, torch.compile, DDP (multi‑GPU), synthetic DG‑Market dataset, adversarial training, etc.
Inference / testing	TensorRT, PyTorch JIT, Conv‑BN fusion, multiple‑query evaluation, GPU‑accelerated re‑ranking, visualisation of rankings and heat‑maps.
Resources	8‑minute tutorial, Chinese video, Colab notebook for free‑GPU training, pre‑trained model downloads.
Licensing	MIT License.
Community	Over 2 500 citations, several thousand GitHub stars, and a history of updates up to 2026 (e.g., DinoV3 support).

---

How to get started (quick‑start)

1. Install – clone the repo and install the required Python packages (PyTorch ≥ 1.8, plus the listed dependencies).
2. Prepare data – follow the Dataset Preparation section to download and format a re‑ID dataset such as Market‑1501.
3. Train – run a single command, e.g.:

   python train.py --train_all               # vanilla ResNet‑50
   python train.py --use_swin --name swin   # Swin‑Transformer model
   

   Add flags like --fp16, --circle, --DG, --usam to enable specific tricks.
4. Test / evaluate – after training, evaluate with:

   python test.py --name <model_name>
   python evaluate_gpu.py               # fast GPU re‑ranking
   

5. Visualise – use the provided demo scripts to see ranking results or heat‑maps.

---

Who might use this?

• Researchers needing a reliable baseline to compare new re‑ID ideas against.
• Practitioners who want a ready‑made, memory‑efficient model for deployment (e.g., on edge GPUs).
• Students looking for a tutorial‑style code base that demonstrates common computer‑vision training tricks.

---

Why it matters

Person re‑identification is a core component of many surveillance, retail analytics, and autonomous‑vehicle perception pipelines. By bundling a wide range of modern architectures, loss functions, and performance‑optimising tricks into a single, well‑documented repository, this project lowers the barrier to entry and speeds up experimentation.

---

Citation

If you use the code or the pre‑trained models in a publication, cite the repository as indicated in the README (the authors provide a ready‑to‑copy BibTeX entry).