Heedless Backbones

NAT 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

---------

Cityscapes (val)

Cityscapes (test)

ADE20K (val)

ADE20K (test)

PASCAL VOC 2007 (val)

PASCAL VOC 2007 (test)

metric

Select an option

mIoU^ms

pAcc^ms

mAcc^ms

mIoU^ss

pAcc^ss

mAcc^ss

GFLOPs

head

Select an option

---------

UPerNet

Mask2Former

Panoptic FPN

SETR

resolution

Select an option

----------

512x2048

640x2560

axis

Select an option

Results

Parameters (M)

Images / Second

GFLOPs

Publication Date

Filters

pretrain dataset

Select an option

---------

ImageNet-1k

ImageNet-22k

JFT-300M

JFT-3B

MegData73M

pretrain method

Select an option

----------

Supervised

Sup. + TL

FCMAE

MAE

CL

MAP

Legend

Legend attribute

Select an option

Family

Pretrain Dataset

Semantic Segmentation Head

Semantic Segmentation Resolution

Semantic Segmentation Training Epochs

Legend attribute (second)

Select an option

----------

Family

Pretrain Method

Semantic Segmentation Head

Semantic Segmentation Resolution

Semantic Segmentation Training Epochs

model	params (m)	pretrain	head	train	GFLOPs	mIoU^ms
NAT-M	20.0	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	900.0	46.4
NAT-T	28.0	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	934.0	48.4
NAT-S	51.0	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	1010.0	49.5
NAT-B	90.0	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	1137.0	49.7

model	params (m)	pretrain	finetune	gflops	IN-1k
NAT-M	20.0	IN-1k : Sup. : 300	— : — : —	2.7	81.8/—
NAT-T	28.0	IN-1k : Sup. : 300	— : — : —	4.3	83.2/—
NAT-S	51.0	IN-1k : Sup. : 300	— : — : —	7.8	83.7/—
NAT-B	90.0	IN-1k : Sup. : 300	— : — : —	13.7	84.3/—

COCO (val)

model	pretrain	head	train	gflops	mAP^b	AP^b₅₀	AP^b₇₅	mAP^b_s	mAP^b_m	mAP^b_l
NAT-M	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	225.0	46.5	68.1	51.3	—	—	—
NAT-M	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	704.0	50.3	68.9	54.9	—	—	—
NAT-T	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	258.0	47.7	69.0	52.6	—	—	—
NAT-T	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	737.0	51.4	70.0	55.9	—	—	—
NAT-S	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	330.0	48.4	69.8	53.2	—	—	—
NAT-S	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	809.0	52.0	70.4	56.3	—	—	—
NAT-B	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	931.0	52.5	71.1	57.1	—	—	—

COCO (val)

model	pretrain	head	train	gflops	mAP^m	AP^m₅₀	AP^m₇₅	mAP^m_s	mAP^m_m	mAP^m_l
NAT-M	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	225.0	41.7	65.2	44.7	—	—	—
NAT-M	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	704.0	43.6	66.4	47.2	—	—	—
NAT-T	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	258.0	42.6	66.1	45.9	—	—	—
NAT-T	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	737.0	44.5	67.6	47.9	—	—	—
NAT-S	IN-1k : Sup. : 300	Mask R-CNN	COCO (train) : 36	330.0	43.2	66.9	46.5	—	—	—
NAT-S	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	809.0	44.9	68.1	48.6	—	—	—
NAT-B	IN-1k : Sup. : 300	Cascade Mask R-CNN	COCO (train) : 36	931.0	45.2	68.6	49.0	—	—	—

ADE20K (val)

model	pretrain	head	train	gflops	mIoU^ms	pAcc^ms	mAcc^ms	mIoU^ss	pAcc^ss	mAcc^ss
NAT-M	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	900.0	46.4	—	—	45.1	—	—
NAT-T	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	934.0	48.4	—	—	47.1	—	—
NAT-S	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	1010.0	49.5	—	—	48.0	—	—
NAT-B	IN-1k : Sup. : 300	UPerNet	ADE20K (train) : 128 : 512	1137.0	49.7	—	—	48.5	—	—