A 2GB-GPU ReID baseline cited over 2,500 times

What it does

This repo is a long-running PyTorch baseline for object re-identification—mostly person and vehicle ReID—meant to give researchers a common, compact starting point. It wires together standard backbones, a dozen loss functions, and training tricks like random erasing and warm-up into one tweakable codebase. Think of it as a reference implementation you can argue with.

The interesting bit

Its staying power comes from relentless maintenance: the authors keep folding in new backbones (now including DinoV3) and inference optimizations like TensorRT and GPU re-ranking, while keeping the training footprint small enough for 2GB GPUs using native mixed precision. That combination of longevity and thrift is rare in computer-vision baselines.

Key highlights

• Native bf16/fp16 support means training fits on 2GB of VRAM without apex.
• Backbone zoo covers ResNet, ResNet-ibn, DenseNet, Swin, EfficientNet, HRNet, and DinoV3.
• Loss functions include Circle, Triplet, ArcFace, CosFace, Contrastive, Sphere, Lifted, and Instance losses.
• Inference shortcuts: TensorRT, PyTorch JIT, Conv-BN fusion, and a GPU re-ranking mode.
• Ships with an 8-minute tutorial and training-curve visualization for the impatient.

Caveats

• The authors warn that some provided hyperparameters are “far from optimal,” so reproducing quoted numbers will require tuning.

Verdict

Grab it if you need a lightweight, well-trodden ReID benchmark to test a new loss or backbone. Look elsewhere if you want a drop-in production tracker or a zero-config state-of-the-art model.