MODEL_ZOO.md

Model Zoo

For the leaderboard on public benchmarks, please refer to LEADERBOARD.md.

Note: all the models below are selected by the performance on validation sets.

Unsupervised learning (USL) on object re-ID

• Direct infer models are directly tested on the re-ID datasets with ImageNet pre-trained weights.

Market-1501

Method	Backbone	Pre-trained	mAP(%)	top-1(%)	top-5(%)	top-10(%)	Train time	Download
Direct infer	ResNet50	ImageNet	2.2	6.7	14.9	20.1	n/a
UDA_TP	ResNet50	ImageNet	34.7	58.6	74.0	78.9	~2h	[config] [model]
strong_baseline	ResNet50	ImageNet	70.5	87.9	95.7	97.1	~2.5h	[config] [model]
MMT	ResNet50	ImageNet	74.3	88.1	96.0	97.5	~4.5h	[config] [model]
SpCL	ResNet50	ImageNet	76.0	89.5	96.2	97.5	~2h	[config] [model]

DukeMTMC-reID

Method	Backbone	Pre-trained	mAP(%)	top-1(%)	top-5(%)	top-10(%)	Train time	Download
Direct infer	ResNet50	ImageNet	2.3	7.5	14.7	18.1	n/a
UDA_TP	ResNet50	ImageNet	42.3	64.4	76.0	79.9	~2h	[config] [model]
strong_baseline	ResNet50	ImageNet	54.7	72.9	83.5	87.2	~2.5h	[config] [model]
MMT	ResNet50	ImageNet	60.3	75.6	86.0	89.2	~4.5h	[config] [model]
SpCL	ResNet50	ImageNet	67.1	82.4	90.8	93.0	~2h	[config] [model]

... (TBD)

Unsupervised domain adaptation (UDA) on object re-ID

• Direct infer models are trained on the source-domain datasets (source_pretrain) and directly tested on the target-domain datasets.
• UDA methods (MMT, SpCL, etc.) starting from ImageNet means that they are trained end-to-end in only one stage without source-domain pre-training.

DukeMTMC-reID -> Market-1501

Method	Backbone	Pre-trained	mAP(%)	top-1(%)	top-5(%)	top-10(%)	Train time	Download
Direct infer	ResNet50	DukeMTMC	27.2	58.9	75.7	81.4	~1h	[config] [model]
UDA_TP	ResNet50	DukeMTMC	52.3	76.0	87.8	91.9	~2h	[config] [model]
strong_baseline	ResNet50	ImageNet	75.6	90.9	96.6	97.8	~3h	[config] [model]
MMT	ResNet50	ImageNet	80.9	92.2	97.6	98.4	~6h	[config] [model]
SpCL	ResNet50	ImageNet	78.2	90.5	96.6	97.8	~3h	[config] [model]

Market-1501 -> DukeMTMC-reID

Method	Backbone	Pre-trained	mAP(%)	top-1(%)	top-5(%)	top-10(%)	Train time	Download
Direct infer	ResNet50	Market	28.1	49.3	64.3	69.7	~1h	[config] [model]
UDA_TP	ResNet50	Market	45.7	65.5	78.0	81.7	~2h	[config] [model]
strong_baseline	ResNet50	ImageNet	60.4	75.9	86.2	89.8	~3h	[config] [model]
MMT	ResNet50	ImageNet	67.7	80.3	89.9	92.9	~6h	[config] [model]
SpCL	ResNet50	ImageNet	70.4	83.8	91.2	93.4	~3h	[config] [model]

... (TBD)