SOTA Vision Foundation Models Leaderboard (08/Sep/2024)

Table of Contents - Overview - Image Classification Leaderboard - Object Detection Leaderboard - Instance Segmentation Leaderboard

Overview

Ranking of Vision Foundation Models by Task.

• Last update: (26/Sep/2024)

Image Classification Leaderboard

Rank	Model Name	Architecture	Pretraining Data	Accuracy	Parameters	Datasets	Release Date
1	ViT-G/14	Vision Transformer (ViT)	JFT-3B	90.3%	6.5B	ImageNet-1k, ImageNet-21k	2022
2	Swin Transformer	Hierarchical Transformer	ImageNet-22k	88.4%	197M	ImageNet-1k, COCO	2021
3	ConvNeXt	Convolution-based Model	ImageNet-22k	87.8%	350M	ImageNet-1k	2022
4	DeiT-III	Data-efficient ViT	ImageNet-1k	87.1%	87M	ImageNet-1k	2021
5	EfficientNetV2	EfficientNet Architecture	ImageNet-21k	85.7%	480M	ImageNet-1k	2021
6	ResNet-RS	Residual Networks	ImageNet-1k	85.2%	100M	ImageNet-1k	2021
7	CoAtNet	Convolution-Attention Hybrid	ImageNet-22k	84.9%	366M	ImageNet-1k	2021
8	NFNet-F6	Normalizer-Free Network	ImageNet-1k	86.5%	438M	ImageNet-1k	2021
9	MLP-Mixer	MLP-based Vision Model	ImageNet-1k	85.3%	59M	ImageNet-1k	2021
10	RegNetY	Regularization Networks	ImageNet-1k	85.4%	145M	ImageNet-1k	2021

---

Object Detection Leaderboard

Rank	Model Name	Architecture	Backbone	mAP (COCO)	Params	GFLOPs	Datasets	Release Date
1	Swin-L	Swin Transformer	Swin-L	58.7	284M	1382	COCO	2021
2	DETR	End-to-End Detection Transformer	ResNet-50	48.1	41M	86	COCO	2020
3	YOLOv7	You Only Look Once	CSP-Darknet53	56.8	67M	18	COCO	2022
4	EfficientDet-D7	EfficientNet Backbone	EfficientNet	52.2	51M	325	COCO	2020
5	Cascade Mask R-CNN	CNN + Cascade	ResNet-101	53.3	101M	390	COCO	2019
6	Faster R-CNN	CNN + Region Proposal	ResNet-50	42.1	41M	184	COCO	2017
7	CenterNet	Center-based Detection	Hourglass-104	47.0	202M	630	COCO	2020
8	FCOS	Fully Convolutional Detection	ResNeXt-101	45.2	64M	128	COCO	2019
9	YOLOv4	You Only Look Once	CSP-Darknet53	51.5	65M	117	COCO	2020
10	RetinaNet	Focal Loss-based Detection	ResNet-101	40.1	56M	96	COCO	2018

---

Instance Segmentation Leaderboard

Rank	Model Name	Architecture	Backbone	AP (COCO)	Params	GFLOPs	Datasets	Release Date
1	Swin-L Mask R-CNN	Swin Transformer	Swin-L	53.8	284M	1382	COCO	2021
2	Cascade Mask R-CNN	Cascade CNN	ResNet-101	50.2	101M	390	COCO	2019
3	SOLOv2	Segment Objects by Locations	ResNet-50	42.1	34M	113	COCO	2020
4	Mask R-CNN	Region Proposal + CNN	ResNet-101	39.8	101M	296	COCO	2017
5	PointRend	Point-based Segmentation	ResNet-50	41.2	46M	96	COCO	2020
6	HTC	Hybrid Task Cascade	ResNet-101	43.6	128M	400	COCO	2019
7	YOLACT++	Real-time Instance Segmentation	ResNet-101	31.2	74M	100	COCO	2019
8	CondInst	Conditional Instance Segmentation	ResNet-50	39.1	31M	73	COCO	2020
9	TensorMask	Mask Head CNN	ResNet-101	37.1	130M	430	COCO	2019
10	DeepLabV3+	Atrous CNN for Segmentation	Xception-71	35.4	43M	100	COCO	2018