Swin Family

Paper | GitHub

axis

Select an option

Results

Parameters (M)

Images / Second

Publication Date

task

Select an option

---------

Object Detection

Instance Segmentation

Classification

Semantic Segmentation

Panoptic Segmentation

dataset

Select an option

---------

ImageNet-1k

ImageNet-A

ImageNet-R

ImageNet-Sketch

ImageNet-C

ImageNet-C-bar

ImageNet-V2

ImageNet-ReaL

metric

Select an option

----------

Top-1

Top-5

GFLOPs

resolution

Select an option

----------

224x224

384x384

512x512

axis

Select an option

Results

Parameters (M)

Images / Second

GFLOPs

Publication Date

Filters

pretrain dataset

Select an option

---------

JFT-3B

ImageNet-1k

ImageNet-22k

JFT-300M

pretrain method

Select an option

----------

Supervised

FCMAE

MAE

Legend

Legend attribute

Select an option

Family

Pretrain Dataset

Pretrain Method

Classification Resolution

Legend attribute (second)

Select an option

----------

Pretrain Dataset

Pretrain Method

Classification Resolution

model	params (m)	pretrain	finetune	GFLOPs	Top-1
Swin-T	29.0	IN-1k : Sup. : 300	— : — : —	4.5	81.3
Swin-T	29.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	4.5	80.9
Swin-B	88.0	IN-1k : Sup. : 300	— : — : —	15.4	83.5
Swin-B	88.0	IN-1k : Sup. : 300	IN-1k : 30 : 384	47.0	84.5
Swin-B	88.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	15.4	85.2
Swin-B	88.0	IN-22k : Sup. : 90	IN-1k : 30 : 384	47.0	86.4
Swin-S	50.0	IN-1k : Sup. : 300	— : — : —	8.7	83.2
Swin-S	50.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	8.7	83.2
Swin-L	197.0	IN-22k : Sup. : 90	IN-1k : 30 : 384	103.9	87.3
Swin-L	197.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	34.5	86.3

model	params (m)	pretrain	finetune	gflops	IN-1k	IN-C↓	IN-A	IN-R	IN-Sketch	IN-V2
Swin-T	29.0	IN-1k : Sup. : 300	— : — : —	4.5	81.3/95.5	62.0/—	21.6/—	41.3/—	29.1/—	69.5/—
Swin-T	29.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	4.5	80.9/96.0	—/—	—/—	—/—	—/—	—/—
Swin-B	88.0	IN-1k : Sup. : 300	— : — : —	15.4	83.5/96.5	54.4/—	35.8/—	46.6/—	32.4/—	—/—
Swin-B	88.0	IN-1k : Sup. : 300	IN-1k : 30 : 384	47.0	84.5/97.0	—/—	—/—	—/—	—/—	—/—
Swin-B	88.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	15.4	85.2/97.5	—/—	—/—	—/—	—/—	74.6/—
Swin-B	88.0	IN-22k : Sup. : 90	IN-1k : 30 : 384	47.0	86.4/98.0	—/—	—/—	—/—	—/—	76.3/—
Swin-S	50.0	IN-1k : Sup. : 300	— : — : —	8.7	83.2/96.2	—/—	—/—	—/—	—/—	71.8/—
Swin-S	50.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	8.7	83.2/97.0	—/—	—/—	—/—	—/—	—/—
Swin-L	197.0	IN-22k : Sup. : 90	IN-1k : 30 : 384	103.9	87.3/98.2	—/—	—/—	—/—	—/—	77.0/—
Swin-L	197.0	IN-22k : Sup. : 90	IN-1k : 30 : 224	34.5	86.3/97.9	—/—	—/—	—/—	—/—	76.3/—

COCO (test)

model	pretrain	head	train	gflops	mAP^b	AP^b₅₀	AP^b₇₅	mAP^b_s	mAP^b_m	mAP^b_l
Swin-L	IN-22k : Sup. : 90	HTC++	COCO (train) : 72	1470.0	57.7	—	—	—	—	—

COCO (val)

model	pretrain	head	train	gflops	mAP^b	AP^b₅₀	AP^b₇₅	mAP^b_s	mAP^b_m	mAP^b_l
Swin-T	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	267.0	46.0	68.1	50.3	—	—	—
Swin-T	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	745.0	50.4	69.2	54.7	—	—	—
Swin-B	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	982.0	51.9	70.5	56.4	—	—	—
Swin-B	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 50	700.0	50.1	—	—	—	—	—
Swin-B	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 50	982.0	52.7	—	—	—	—	—
Swin-B	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 36	982.0	53.0	71.8	57.5	—	—	—
Swin-B	IN-22k : Sup. : 90	Mask R-CNN	COCO (train) : 50	700.0	51.4	—	—	—	—	—
Swin-B	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 50	982.0	54.0	—	—	—	—	—
Swin-B	IN-22k : Sup. : 90	HTC++	COCO (train) : 72	1043.0	56.4	—	—	—	—	—
Swin-S	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	359.0	48.5	—	—	—	—	—
Swin-S	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	838.0	51.9	70.7	56.3	—	—	—
Swin-L	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 36	1382.0	53.9	72.4	58.8	—	—	—
Swin-L	IN-22k : Sup. : 90	Mask R-CNN	COCO (train) : 50	1100.0	52.4	—	—	—	—	—
Swin-L	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 50	1382.0	54.8	—	—	—	—	—
Swin-L	IN-22k : Sup. : 90	HTC++	COCO (train) : 72	1470.0	57.1	—	—	—	—	—

COCO (test)

model	pretrain	head	train	gflops	mAP^m	AP^m₅₀	AP^m₇₅	mAP^m_s	mAP^m_m	mAP^m_l
Swin-L	IN-22k : Sup. : 90	HTC++	COCO (train) : 72	1470.0	50.2	—	—	—	—	—

COCO (val)

model	pretrain	head	train	gflops	mAP^m	AP^m₅₀	AP^m₇₅	mAP^m_s	mAP^m_m	mAP^m_l
Swin-T	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	267.0	41.6	65.1	44.9	—	—	—
Swin-T	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	745.0	43.7	66.6	47.3	—	—	—
Swin-B	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	982.0	45.0	68.1	48.9	—	—	—
Swin-B	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 50	600.0	44.5	—	—	—	—	—
Swin-B	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 50	982.0	45.5	—	—	—	—	—
Swin-B	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 36	982.0	45.8	69.4	49.7	—	—	—
Swin-B	IN-22k : Sup. : 90	Mask R-CNN	COCO (train) : 50	700.0	45.4	—	—	—	—	—
Swin-B	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 50	982.0	46.5	—	—	—	—	—
Swin-B	IN-22k : Sup. : 90	HTC++	COCO (train) : 72	1043.0	49.1	—	—	—	—	—
Swin-S	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	359.0	43.3	—	—	—	—	—
Swin-S	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	838.0	45.0	68.2	48.8	—	—	—
Swin-L	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 36	1382.0	46.7	70.1	50.8	—	—	—
Swin-L	IN-22k : Sup. : 90	Mask R-CNN	COCO (train) : 50	1100.0	46.2	—	—	—	—	—
Swin-L	IN-22k : Sup. : 90	Cascade Mask R-CNN	COCO (train) : 50	1382.0	47.3	—	—	—	—	—
Swin-L	IN-22k : Sup. : 90	HTC++	COCO (train) : 72	1470.0	49.5	—	—	—	—	—

ADE20K (val)

model	pretrain	head	train	gflops	mIoU^ms	pAcc^ms	mAcc^ms	mIoU^ss	pAcc^ss	mAcc^ss
Swin-T	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	945.0	45.8	—	—	44.5	—	—
Swin-T	IN-1k : Sup. : 300	Panoptic FPN	ADE20K (train) : 64 : 512	182.0	—	—	—	41.5	—	—
Swin-B	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	1188.0	49.7	—	—	48.1	—	—
Swin-B	IN-1k : Sup. : 300	Panoptic FPN	ADE20K (train) : 64 : 512	422.0	—	—	—	46.0	—	—
Swin-B	IN-22k : Sup. : 90	UPerNet	ADE20K (train) : 128 : 640	1841.0	51.7	—	—	50.0	—	—
Swin-S	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	1038.0	49.5	—	—	47.6	—	—
Swin-S	IN-1k : Sup. : 300	Panoptic FPN	ADE20K (train) : 64 : 512	274.0	—	—	—	45.2	—	—
Swin-L	IN-22k : Sup. : 90	UPerNet	ADE20K (train) : 128 : 640	3230.0	53.5	—	—	52.1	—	—

ADE20K (test)

model	pretrain	head	train	gflops	mIoU^ms	pAcc^ms	mAcc^ms	mIoU^ss	pAcc^ss	mAcc^ss
Swin-L	IN-22k : Sup. : 90	UPerNet	ADE20K (train) : 128 : 640	3230.0	62.8	—	—	—	—	—