MODEL_ZOO.md

FastReID Model Zoo and Baselines

Introduction

This file documents collection of baselines trained with fastreid. All numbers were obtained with 1 NVIDIA V100 GPU. The software in use were PyTorch 1.6, CUDA 10.1.

In addition to these official baseline models, you can find more models in projects/.

How to Read the Tables

• The "Name" column contains a link to the config file. Running tools/train_net.py with this config file and 1 GPU will reproduce the model.

Common Settings for all Person reid models

BoT:

Bag of Tricks and A Strong Baseline for Deep Person Re-identification. CVPRW2019, Oral.

AGW:

ReID-Survey with a Powerful AGW Baseline.

MGN:

Learning Discriminative Features with Multiple Granularities for Person Re-Identification

SBS:

stronger baseline on top of BoT:

Bag of Freebies(BoF):

1. Circle loss
2. Freeze backbone training
3. Cutout data augmentation & Auto Augmentation
4. Cosine annealing learning rate decay
5. Soft margin triplet loss

Bag of Specials(BoS):

1. Non-local block
2. GeM pooling

Market1501 Baselines

BoT:

Method	Pretrained	Rank@1	mAP	mINP	download
BoT(R50)	ImageNet	94.4%	86.1%	59.4%	model
BoT(R50-ibn)	ImageNet	94.9%	87.6%	64.1%	model
BoT(S50)	ImageNet	95.2%	88.7%	66.9%	model
BoT(R101-ibn)	ImageNet	95.4%	88.9%	67.4%	model

AGW:

Method	Pretrained	Rank@1	mAP	mINP	download
AGW(R50)	ImageNet	95.3%	88.2%	66.3%	model
AGW(R50-ibn)	ImageNet	95.1%	88.7%	67.1%	model
AGW(S50)	ImageNet	95.3%	89.3%	68.5%	model
AGW(R101-ibn)	ImageNet	95.5%	89.5%	69.5%	model

SBS:

Method	Pretrained	Rank@1	mAP	mINP	download
SBS(R50)	ImageNet	95.4%	88.2%	64.8%	model
SBS(R50-ibn)	ImageNet	95.7%	89.3%	67.5%	model
SBS(S50)	ImageNet	95.8%	89.4%	67.6%	model
SBS(R101-ibn)	ImageNet	96.3%	90.3%	70.0%	model

MGN:

Method	Pretrained	Rank@1	mAP	mINP	download
SBS(R50-ibn)	ImageNet	95.8%	89.8%	67.7%	model

DukeMTMC Baseline

BoT:

Method	Pretrained	Rank@1	mAP	mINP	download
BoT(R50)	ImageNet	87.2%	77.0%	42.1%	model
BoT(R50-ibn)	ImageNet	89.3%	79.6%	45.2%	model
BoT(S50)	ImageNet	90.0%	80.13%	45.8%	model
BoT(R101-ibn)	ImageNet	91.2%	81.2%	47.5%	model

AGW:

Method	Pretrained	Rank@1	mAP	mINP	download
AGW(R50)	ImageNet	89.0%	79.9%	46.1%	model
AGW(R50-ibn)	ImageNet	90.5%	80.8%	47.6%	model
AGW(S50)	ImageNet	90.9%	82.4%	49.2%	model
AGW(R101-ibn)	ImageNet	91.7%	82.3%	50.0%	model

SBS:

Method	Pretrained	Rank@1	mAP	mINP	download
SBS(R50)	ImageNet	90.3%	80.3%	46.5%	model
SBS(R50-ibn)	ImageNet	90.8%	81.2%	47.0%	model
SBS(S50)	ImageNet	91.0%	81.4%	47.6%	model
SBS(R101-ibn)	ImageNet	91.9%	83.6%	51.5%	model

MGN:

Method	Pretrained	Rank@1	mAP	mINP	download
SBS(R50-ibn)	ImageNet	91.1%	82.0%	46.8%	model

MSMT17 Baseline

BoT:

Method	Pretrained	Rank@1	mAP	mINP	download
BoT(R50)	ImageNet	74.1%	50.2%	10.4%	model
BoT(R50-ibn)	ImageNet	77.0%	54.4%	12.5%	model
BoT(S50)	ImageNet	80.8%	59.9%	16.3%	model
BoT(R101-ibn)	ImageNet	81.0%	59.4%	15.6%	model

AGW:

Method	Pretrained	Rank@1	mAP	mINP	download
AGW(R50)	ImageNet	78.3%	55.6%	12.9%	model
AGW(R50-ibn)	ImageNet	81.2%	59.7%	15.3%	model
AGW(S50)	ImageNet	82.6%	62.6%	17.7%	model
AGW(R101-ibn)	ImageNet	82.0%	61.4%	17.3%	model

SBS:

Method	Pretrained	Rank@1	mAP	mINP	download
SBS(R50)	ImageNet	81.8%	58.4%	13.9%	model
SBS(R50-ibn)	ImageNet	83.9%	60.6%	15.2%	model
SBS(S50)	ImageNet	84.1%	61.7%	15.2%	model
SBS(R101-ibn)	ImageNet	84.8%	62.8%	16.3%	model

MGN:

Method	Pretrained	Rank@1	mAP	mINP	download
SBS(R50-ibn)	ImageNet	85.1%	65.4%	18.4%	-

VeRi Baseline

SBS:

Method	Pretrained	Rank@1	mAP	mINP	download
SBS(R50-ibn)	ImageNet	97.0%	81.9%	46.3%	model

VehicleID Baseline

BoT:
Test protocol: 10-fold cross-validation; trained on 4 NVIDIA P40 GPU.

Method	Pretrained	Testset size						download
		Small		Medium		Large
		Rank@1	Rank@5	Rank@1	Rank@5	Rank@1	Rank@5
BoT(R50-ibn)	ImageNet	86.6%	97.9%	82.9%	96.0%	80.6%	93.9%	model

VERI-Wild Baseline

BoT:
Test protocol: Trained on 4 NVIDIA P40 GPU.

Method	Pretrained	Testset size									download
		Small			Medium			Large
		Rank@1	mAP	mINP	Rank@1	mAP	mINP	Rank@1	mAP	mINP
BoT(R50-ibn)	ImageNet	96.4%	87.7%	69.2%	95.1%	83.5%	61.2%	92.5%	77.3%	49.8%	model