How FastReID V1.0 Revolutionizes General Object Re‑Identification

In June 2020 the JD AI Research Institute released FastReID, a PyTorch‑based open‑source library for general object re‑identification (ReID). After continuous optimization, FastReID V1.0 was announced on January 18, 2021 and has since been widely applied within JD for smart campuses, smart buildings, supply‑chain logistics, and offline retail.

ReID, also known as cross‑camera tracking or re‑identification, retrieves a specific target (person, vehicle, etc.) from a set of candidate images, enabling continuous tracking across multiple cameras and enhancing data continuity while reducing manual labor.

Typical applications include mask detection for epidemic control, license‑plate monitoring on highways, and personalized services in new‑retail scenarios, though challenges such as occlusion and privacy remain.

FastReID adopts the design philosophy of Detectron2 and provides a highly modular and extensible architecture. The training pipeline consists of image preprocessing, backbone, aggregation module, head, training strategies, and loss functions. The inference pipeline includes metric computation and post‑processing methods such as K‑reciprocal coding and query expansion.

The V1.0 update introduces faster distributed training, one‑click model export to Caffe/ONNX/TensorRT, model distillation, automatic hyper‑parameter search, and support for additional tasks like person attribute recognition, landmark retrieval, tracking, and image classification.

Model Distillation : By applying KL divergence and symmetric JS divergence losses, FastReID improves accuracy of smaller student models without altering their architecture. Feature distillation is achieved via overhaul‑distillation, which modifies the backbone only within a separate project to avoid affecting existing models. Users can enable loss‑only distillation with the Distiller meta‑architecture or loss + feature distillation with DistillerOverhaul, specifying teacher configuration and weights.

# teacher model training
python3 projects/FastDistill/train_net.py \
    --config-file projects/FastDistill/configs/sbs_r101ibn.yml \
    --num-gpus 4

# loss distillation
python3 projects/FastDistill/train_net.py \
    --config-file projects/FastDistill/configs/kd-sbs_r101ibn-sbs_r34.yaml \
    --num-gpus 4 \
    MODEL.META_ARCHITECTURE Distiller \
    KD.MODEL_CONFIG projects/FastDistill/logs/dukemtmc/r101_ibn/config.yaml \
    KD.MODEL_WEIGHTS projects/FastDistill/logs/dukemtmc/r101_ibn/model_best.pth

# loss + overhaul distillation
python3 projects/FastDistill/train_net.py \
    --config-file projects/FastDistill/configs/kd-sbs_r101ibn-sbs_r34.yaml \
    --num-gpus 4 \
    MODEL.META_ARCHITECTURE DistillerOverhaul \
    KD.MODEL_CONFIG projects/FastDistill/logs/dukemtmc/r101_ibn/config.yaml \
    KD.MODEL_WEIGHTS projects/FastDistill/logs/dukemtmc/r101_ibn/model_best.pth

Automatic Hyper‑Parameter Search : FastReID V1.0 integrates Bayesian optimization (BOHB) for hyper‑parameter tuning. Users can launch distributed trials with a single command, e.g.:

python3 projects/FastTune/tune_net.py \
    --config-file projects/FastTune/configs/search_trial.yml \
    --num-trials 12 --srch-alog "bohb"

Since its open‑source debut, FastReID has achieved state‑of‑the‑art results on various benchmarks and serves as the core engine for JD’s internal ReID research and development. Future plans include adding more advanced algorithms, improving documentation, and further enhancing scalability and stability for large‑scale data.