notesum.ai

Published at November 21

cs.CV

cs.AI

Released Date: November 21, 2024

Authors: Lei Zhu¹, Xinjiang Wang², Wayne Zhang², Rynson W. H. Lau

Aff.: ¹City University of Hong Kong; ²Sensetime Research

Method		FLOPs	#Param.	Top-1
Standard supervised training
Swin-T [35]	[ICCV’21]	4.5G	29M	81.3
PVT-S [51]	[ICCV’21]	3.8G	25M	79.8
CSWin-T [14]	[CVPR’22]	4.5G	23M	82.7
CMT-S [18]	[CVPR’22]	4.0G	25M	83.5
RegionViT-S [3]	[ICLR’22]	5.7G	31M	83.3
CrossFormer-S [53]	[ICLR’23]	5.3G	31M	82.5
MaxViT-T [47]	[ECCV’22]	5.6G	31M	83.6
MOAT-0 [59]	[ICLR’23]	5.7G	28M	83.3
NAT-T [20]	[CVPR’23]	4.3G	28M	83.2
InternImage-T [52]	[CVPR’23]	5G	30M	83.5
Flatten-T [19]	[ICCV’23]	4.3G	21M	83.1
SG-Former-S [43]	[ICCV’23]	4.8G	23M	83.2
GLNet-4G (ours)		4.5G	27M	83.7
Swin-S [35]	[ICCV’21]	8.7G	50M	83.0
CSwin-S [14]	[CVPR’22]	6.9G	35M	83.6
RegionViT-M [3]	[ICLR’22]	7.9G	42M	83.4
MaxViT-S [47]	[ECCV’23]	11.7G	69M	84.4
MOAT-1 [59]	[ICLR’23]	9.1G	42M	84.2
NAT-S [20]	[CVPR’23]	7.8G	51M	83.7
InternImage-S [52]	[CVPR’23]	8G	50M	84.2
BiFormer-B [70]	[CVPR’23]	9.8G	57M	84.3
Flatten-S [19]	[ICCV’23]	6.9G	35M	83.8
SG-Former-M [43]	[ICCV’23]	7.5G	39M	84.1
SMT-B [34]	[ICCV’23]	7.7G	32M	84.3
GLNet-9G (ours)		9.7G	61M	84.5