[Paper Review] FastReID: A Pytorch Toolbox for General Instance Re-identification

Lingxiao He, Xingyu Liao+4 more|arXiv (Cornell University)|Jun 4, 2020

Video Surveillance and Tracking Methods44 references123 citations

TL;DR

FastReID presents a modular PyTorch toolbox for general instance re-identification, with scalable training, rich evaluation metrics, and deployment tools, achieving strong results on standard ReID benchmarks.

ABSTRACT

General Instance Re-identification is a very important task in the computer vision, which can be widely used in many practical applications, such as person/vehicle re-identification, face recognition, wildlife protection, commodity tracing, and snapshop, etc.. To meet the increasing application demand for general instance re-identification, we present FastReID as a widely used software system in JD AI Research. In FastReID, highly modular and extensible design makes it easy for the researcher to achieve new research ideas. Friendly manageable system configuration and engineering deployment functions allow practitioners to quickly deploy models into productions. We have implemented some state-of-the-art projects, including person re-id, partial re-id, cross-domain re-id and vehicle re-id, and plan to release these pre-trained models on multiple benchmark datasets. FastReID is by far the most general and high-performance toolbox that supports single and multiple GPU servers, you can reproduce our project results very easily and are very welcome to use it, the code and models are available at https://github.com/JDAI-CV/fast-reid.

Motivation & Objective

• Motivate a unified, extensible library for general instance re-identification to bridge research and production needs.
• Provide modular components (backbone, aggregation, head, losses) to accelerate new ideas without rewriting code.
• Offer comprehensive evaluation and deployment support (multi-GPU training, distillation, model conversion).
• Deliver state-of-the-art pre-trained models for varied ReID tasks (person, cross-domain, partial, vehicle).
• Enable fair, reproducible comparisons across approaches on common benchmarks.

Proposed method

• Proposes a four-module pipeline: image pre-processing, backbone, aggregation, and head.
• Implements multiple backbones (ResNet, ResNeXt, ResNeSt) with attention and IBN variants.
• Supports four aggregation methods (max, average, GeM, attention pooling) and three head types (BN neck, linear, reduction) with batch normalization.
• Employs diverse loss functions (cross-entropy with label smoothing, ArcFace, Circle loss, triplet loss) and training tricks (lr warm-up, backbone freezing, cosine learning rate decay).
• Uses knowledge distillation (teacher-student, logit matching) to obtain lightweight, high-accuracy models.
• Introduces post-processing techniques (K-reciprocal re-ranking, query expansion) and multiple distance metrics (Euclidean, cosine, DSR).
• Provides model conversion tools (PyTorch to Caffe/TensorRT) and a framework for deploying models in production.

Experimental results

Research questions

• RQ1 Can a modular PyTorch toolbox support rapid development and deployment of general instance re-identification models?
• RQ2 What combination of backbones, pooling, and loss functions yields strong performance across multiple ReID tasks (person, cross-domain, partial, vehicle)?
• RQ3 How can training strategies and distillation improve accuracy while enabling lightweight deployment?
• RQ4 What evaluation and post-processing tools best reflect real-world deployment performance for ReID models?

Key findings

• FastReID achieves strong performance on major ReID benchmarks (Market-1501, DukeMTMC, MSMT17) with ResNet101-IBN backbones showing high rank-1/mAP across datasets.
• Incorporating QE (query expansion) and knowledge distillation improves accuracy, delivering higher mAP and rank-1 in reported ablations (e.g., +QE and distillation gains).
• The library provides rich evaluation metrics beyond CMC, including ROC and mINP, and supports detailed post-processing for improved retrieval.
• FastReID offers state-of-the-art pre-trained models for multiple tasks (person, cross-domain, partial, vehicle) and facilitates deployment with model conversion to Caffe and TensorRT.
• The framework demonstrates strong transfer across tasks and domains, enabling practical deployment in production environments.
• table_headers':['Methods','Market1501 R1','Market1501 mAP','DukeMTMC R1','DukeMTMC mAP','MSMT17 R1','MSMT17 mAP'] ,
• table_rows
• [["FastReID (ResNet101-ibn)","96.3","90.3","92.4","83.2","85.1","63.3"],["+ QE (FastReID 101-ibn)","96.5","94.4","93.4","90.1","87.9","76.9"]]