Self-Supervised Image Classification 벤치마크

Self-Supervised Image Classification on ImageNet

288개 결과 · ⬇ CSV · JSON

Top 1 Accuracy

Rank	Model	Top 1 Accuracy	Top 5 Accuracy	Number of Params	Extra Training Data	Paper	Code	Year
1	DINOv2+reg (ViT-g/14)	87.1	–	1100M	✓	Vision Transformers Need Registers	rwightman/pytorch-image-models · facebookresearch/dinov2 · locuslab/massive-activations · +3	2023
2	DINOv2 (ViT-g/14 @448)	86.7%	–	1100M	✓	DINOv2: Learning Robust Visual Features without Supervision	huggingface/transformers · facebookresearch/dinov2 · roboflow/rf-detr · +23	2023
3	DINOv2 (ViT-g/14)	86.5%	–	1100M	✓	DINOv2: Learning Robust Visual Features without Supervision	huggingface/transformers · facebookresearch/dinov2 · roboflow/rf-detr · +23	2023
4	DINOv2 distilled (ViT-L/14)	86.3%	–	307M	✓	DINOv2: Learning Robust Visual Features without Supervision	huggingface/transformers · facebookresearch/dinov2 · roboflow/rf-detr · +23	2023
5	MIM-Refiner (D2V2-ViT-H/14)	84.7%	–	632M		MIM-Refiner: A Contrastive Learning Boost from Intermediate Pre-Trained Representations	ml-jku/MIM-Refiner · BenediktAlkin/vtab1k-pytorch	2024
6	MIM-Refiner (MAE-ViT-2B/14)	84.5%	–	1890M		MIM-Refiner: A Contrastive Learning Boost from Intermediate Pre-Trained Representations	ml-jku/MIM-Refiner · BenediktAlkin/vtab1k-pytorch	2024
6	DINOv2 distilled (ViT-B/14)	84.5%	–	85M	✓	DINOv2: Learning Robust Visual Features without Supervision	huggingface/transformers · facebookresearch/dinov2 · roboflow/rf-detr · +23	2023
8	MIM-Refiner (MAE-ViT-H/14	83.7%	–	632M		MIM-Refiner: A Contrastive Learning Boost from Intermediate Pre-Trained Representations	ml-jku/MIM-Refiner · BenediktAlkin/vtab1k-pytorch	2024
9	MIM-Refiner (D2V2-ViT-L/16)	83.5%	–	307M		MIM-Refiner: A Contrastive Learning Boost from Intermediate Pre-Trained Representations	ml-jku/MIM-Refiner · BenediktAlkin/vtab1k-pytorch	2024
10	MIM-Refiner (MAE-ViT-L/16)	82.8%	–	307M		MIM-Refiner: A Contrastive Learning Boost from Intermediate Pre-Trained Representations	ml-jku/MIM-Refiner · BenediktAlkin/vtab1k-pytorch	2024
11	iBOT (ViT-L/16) (IN22k)	82.3%	–	307M	✓	iBOT: Image BERT Pre-Training with Online Tokenizer	bytedance/ibot · birder/birder	2021
12	MAE-CT (ViT-H/16)	82.2%	–	632M		Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget	ml-jku/mae-ct	2023
13	Mugs (VIT-L/16)	82.1%	–	307M		Mugs: A Multi-Granular Self-Supervised Learning Framework	sail-sg/mugs	2022
14	MAE-CT (ViT-L/16	81.5%	–	307M		Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget	ml-jku/mae-ct	2023
15	EsViT (Swin-B)	81.3	95.5	87M		Efficient Self-supervised Vision Transformers for Representation Learning	microsoft/esvit	2021
15	iBOT (ViT-L/16)	81.3%	–	307M		iBOT: Image BERT Pre-Training with Online Tokenizer	bytedance/ibot · birder/birder	2021
17	DINOv2 distilled (ViT-S/14)	81.1%	–	21M	✓	DINOv2: Learning Robust Visual Features without Supervision	huggingface/transformers · facebookresearch/dinov2 · roboflow/rf-detr · +23	2023
18	MoCo v3 (ViT-BN-L/7)	81.0%	–	304M		An Empirical Study of Training Self-Supervised Vision Transformers	open-mmlab/mmselfsup · facebookresearch/moco-v3 · Westlake-AI/openmixup · +6	2021
19	EsViT(Swin-S)	80.8	–	49M		Efficient Self-supervised Vision Transformers for Representation Learning	microsoft/esvit	2021
20	MSN (ViT-L/7)	80.7%	–	306M		Masked Siamese Networks for Label-Efficient Learning	lightly-ai/lightly · facebookresearch/msn	2022
21	ReLICv2 (ResNet-200 x2)	80.6%	–	250M		Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?	google-deepmind/relicv2	2022
22	MR BarTwins (MR BarTwins)	80.4%	–	–		Masked Reconstruction Contrastive Learning with Information Bottleneck Principle		2022
23	DiGIT	80.3%	–	732M		Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective	DAMO-NLP-SG/DiGIT	2024
23	iBOT-vMF (ViT-B/16)	80.3%	–	85M		DINO as a von Mises-Fisher mixture model		2024
23	DINO (xcit_medium_24_p8)	80.3%	–	84M		Emerging Properties in Self-Supervised Vision Transformers	facebookresearch/dino · lightly-ai/lightly · facebookresearch/vissl · +29	2021
23	PGT (PGT-B w/ Flow)	80.3%	–	70M		Perceptual Group Tokenizer: Building Perception with Iterative Grouping		2023
27	DINO (ViT-B/8)	80.1%	–	80M		Emerging Properties in Self-Supervised Vision Transformers	facebookresearch/dino · lightly-ai/lightly · facebookresearch/vissl · +29	2021
28	SimCLRv2 (ResNet-152 x3, SK)	79.8%	94.9%	795M		Big Self-Supervised Models are Strong Semi-Supervised Learners	google-research/simclr · lightly-ai/lightly · sayakpaul/PAWS-TF · +6	2020
28	SEERv2	79.8%	–	10000M	✓	Vision Models Are More Robust And Fair When Pretrained On Uncurated Images Without Supervision	facebookresearch/vissl	2022
28	PercMAE (ViT-B, dVAE)	79.8%	–	80M	✓	Improving Visual Representation Learning through Perceptual Understanding	tractableai/perceptual-mae	2022
28	ReLICv2 (ResNet200)	79.8%	–	63M		Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?	google-deepmind/relicv2	2022
32	DINO (ViT-S/8)	79.7%	–	21M		Emerging Properties in Self-Supervised Vision Transformers	facebookresearch/dino · lightly-ai/lightly · facebookresearch/vissl · +29	2021
33	BYOL (ResNet-200 x2)	79.6%	94.8%	250M		Bootstrap your own latent: A new approach to self-supervised Learning	deepmind/deepmind-research · alibaba/EasyCV · lucidrains/byol-pytorch · +28	2020
34	ReLICv2 (ResNet-50 4x)	79.4%	–	375M		Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?	google-deepmind/relicv2	2022
35	ReLICv2 (ResNet152)	79.3%	–	58M		Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?	google-deepmind/relicv2	2022
35	BAM (CAFormer-M36)	79.3%	–	–		Unsupervised Representation Learning by Balanced Self Attention Matching	danielshalam/bam	2024
37	MoCo v3 (ViT-BN-H)	79.1%	–	700M		An Empirical Study of Training Self-Supervised Vision Transformers	open-mmlab/mmselfsup · facebookresearch/moco-v3 · Westlake-AI/openmixup · +6	2021
37	Unicom (ViT-B/16)	79.1%	–	80M	✓	Unicom: Universal and Compact Representation Learning for Image Retrieval	OML-Team/open-metric-learning · deepglint/unicom · RocketFlash/easy_metric_learning	2023
39	SMoG (ResNet-50 x4)	79.0%	94.4	375M		Unsupervised Visual Representation Learning by Synchronous Momentum Grouping	lightly-ai/lightly	2022
39	ReLICv2 (ResNet-50 x2)	79%	–	94M		Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?	google-deepmind/relicv2	2022
41	C-BYOL (ResNet-50 2x, 1000 epochs)	78.8%	94.5%	94M		Compressive Visual Representations	google-research/compressive-visual-representations	2021
41	DINO-vMF (ViT-B/16)	78.8%	–	85M		DINO as a von Mises-Fisher mixture model		2024
43	ReLICv2 (ResNet101)	78.7%	–	44M		Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?	google-deepmind/relicv2	2022
44	BYOL (ResNet-50 x4)	78.6%	94.2%	375M		Bootstrap your own latent: A new approach to self-supervised Learning	deepmind/deepmind-research · alibaba/EasyCV · lucidrains/byol-pytorch · +28	2020
45	SwAV (ResNet-50 x5)	78.5%	–	586M		Unsupervised Learning of Visual Features by Contrasting Cluster Assignments	open-mmlab/mmdetection · lightly-ai/lightly · facebookresearch/vissl · +15	2020
46	DINO (ViT-B/16)	78.2%	–	85M		Emerging Properties in Self-Supervised Vision Transformers	facebookresearch/dino · lightly-ai/lightly · facebookresearch/vissl · +29	2021
47	MoCo v3 (ViT-H)	78.1%	–	632M		An Empirical Study of Training Self-Supervised Vision Transformers	open-mmlab/mmselfsup · facebookresearch/moco-v3 · Westlake-AI/openmixup · +6	2021
47	PercMAE (ViT-B)	78.1%	–	80M		Improving Visual Representation Learning through Perceptual Understanding	tractableai/perceptual-mae	2022
47	BAM (ViT-B/16)	78.1%	–	80M		Unsupervised Representation Learning by Balanced Self Attention Matching	danielshalam/bam	2024
50	SMoG (ResNet-50 x2)	78.0%	93.9	94M		Unsupervised Visual Representation Learning by Synchronous Momentum Grouping	lightly-ai/lightly	2022
51	MoCo v3 (ViT-L)	77.6%	–	307M		An Empirical Study of Training Self-Supervised Vision Transformers	open-mmlab/mmselfsup · facebookresearch/moco-v3 · Westlake-AI/openmixup · +6	2021
52	SEER	77.5%	–	1300M	✓	Self-supervised Pretraining of Visual Features in the Wild	facebookresearch/vissl	2021
53	BYOL (ResNet-50 x2)	77.4%	93.6%	94M		Bootstrap your own latent: A new approach to self-supervised Learning	deepmind/deepmind-research · alibaba/EasyCV · lucidrains/byol-pytorch · +28	2020
54	SwAV (ResNet-50 x2)	77.3%	–	94M		Unsupervised Learning of Visual Features by Contrasting Cluster Assignments	open-mmlab/mmdetection · lightly-ai/lightly · facebookresearch/vissl · +15	2020
55	ReLICv2 (ResNet-50)	77.1%	–	25M		Pushing the limits of self-supervised ResNets: Can we outperform supervised learning without labels on ImageNet?	google-deepmind/relicv2	2022
56	DINO (ViT-S/16)	77.0%	–	21M		Emerging Properties in Self-Supervised Vision Transformers	facebookresearch/dino · lightly-ai/lightly · facebookresearch/vissl · +29	2021
56	DINO-vMF (ViT-S/16)	77.0%	–	21M		DINO as a von Mises-Fisher mixture model		2024
58	MoCo v3 (ViT-B/16)	76.7%	–	86M		An Empirical Study of Training Self-Supervised Vision Transformers	open-mmlab/mmselfsup · facebookresearch/moco-v3 · Westlake-AI/openmixup · +6	2021
59	MAE (ViT-H)	76.6%	–	700M		Masked Autoencoders Are Scalable Vision Learners	facebookresearch/mae · lightly-ai/lightly · open-mmlab/mmselfsup · +55	2021
60	SimCLR (ResNet-50 4x)	76.5%	93.2%	375M		A Simple Framework for Contrastive Learning of Visual Representations	tensorflow/models · google-research/simclr · lightly-ai/lightly · +93	2020
61	CoKe (ResNet-50)	76.4%	–	25M		Unsupervised Visual Representation Learning by Online Constrained K-Means	idstcv/coke	2021
61	SMoG (ResNet-50)	76.4%	–	25M		Unsupervised Visual Representation Learning by Synchronous Momentum Grouping	lightly-ai/lightly	2022
63	ReSSL (ResNet-50 w/ Predictor and Stronger Aug)	76.3%	–	24M		Weak Augmentation Guided Relational Self-Supervised Learning	mingkai-zheng/ReSSL	2022
64	ReSSL (ResNet-50 w/ Predictor)	76.0%	–	24M		Weak Augmentation Guided Relational Self-Supervised Learning	mingkai-zheng/ReSSL	2022
65	Triplet (ResNet-50)	75.9%	–	23.56M		Solving Inefficiency of Self-supervised Representation Learning	wanggrun/triplet	2021
66	MAE (ViT-L)	75.8%	–	306M		Masked Autoencoders Are Scalable Vision Learners	facebookresearch/mae · lightly-ai/lightly · open-mmlab/mmselfsup · +55	2021
66	DnC (ResNet-50)	75.8%	–	24M		Divide and Contrast: Self-supervised Learning from Uncurated Data		2021
68	CaCo (ResNet-50)	75.7%	–	24M		CaCo: Both Positive and Negative Samples are Directly Learnable via Cooperative-adversarial Contrastive Learning	maple-research-lab/caco	2022
69	SimCLRv2 (ResNet-50 x2)	75.6%	92.7%	94M		Big Self-Supervised Models are Strong Semi-Supervised Learners	google-research/simclr · lightly-ai/lightly · sayakpaul/PAWS-TF · +6	2020
69	C-BYOL (ResNet-50, 1000 epochs)	75.6%	92.7%	25M		Compressive Visual Representations	google-research/compressive-visual-representations	2021
69	NNCLR (ResNet-50, multi-crop)	75.6%	92.4	25M		With a Little Help from My Friends: Nearest-Neighbor Contrastive Learning of Visual Representations	lightly-ai/lightly · keras-team/keras-io · vturrisi/solo-learn · +1	2021
72	HEXA	75.5%	–	24M		Self-supervised Pre-training with Hard Examples Improves Visual Representations		2020
73	SCE (ResNet-50, multi-crop)	75.4%	–	24M		Similarity Contrastive Estimation for Self-Supervised Soft Contrastive Learning	juliendenize/eztorch · cea-list/sce	2021
74	SwAV (ResNet-50)	75.3%	–	24M		Unsupervised Learning of Visual Features by Contrasting Cluster Assignments	open-mmlab/mmdetection · lightly-ai/lightly · facebookresearch/vissl · +15	2020
74	DINO (ResNet-50)	75.3%	–	24M		Emerging Properties in Self-Supervised Vision Transformers	facebookresearch/dino · lightly-ai/lightly · facebookresearch/vissl · +29	2021
76	InfoMin (ResNeXt-152)	75.2%	–	120M		What Makes for Good Views for Contrastive Learning?	HobbitLong/PyContrast	2020
76	DeepCluster-v2 (ResNet-50)	75.2%	–	24M		Unsupervised Learning of Visual Features by Contrasting Cluster Assignments	open-mmlab/mmdetection · lightly-ai/lightly · facebookresearch/vissl · +15	2020
78	Unicom (ViT-B/32)	75.0%	–	80M	✓	Unicom: Universal and Compact Representation Learning for Image Retrieval	OML-Team/open-metric-learning · deepglint/unicom · RocketFlash/easy_metric_learning	2023
78	MoBY (Swin-T)	75%	–	29M		Self-Supervised Learning with Swin Transformers	microsoft/Swin-Transformer · alibaba/EasyCV · SwinTransformer/Transformer-SSL · +3	2021
80	ReLIC (ResNet-50)	74.8%	–	24M		Representation Learning via Invariant Causal Mechanisms	filipbasara0/relic · filipbasara0/matryoshka-representation-learning	2020
81	ReSSL(ResNet-50) 200ep	74.7%	92.3%	24M		ReSSL: Relational Self-Supervised Learning with Weak Augmentation	vturrisi/solo-learn · KyleZheng1997/ReSSL	2021
81	WCL (ResNet-50)	74.7%	–	24M		Weakly Supervised Contrastive Learning	KyleZheng1997/WCL	2021
83	MV-MR	74.5%	92.1	–		MV-MR: multi-views and multi-representations for self-supervised learning and knowledge distillation	vkinakh/mv-mr	2023
84	FNC (ResNet-50)	74.4%	91.8%	24M		Boosting Contrastive Self-Supervised Learning with False Negative Cancellation	google-research/fnc	2020
85	BYOL (ResNet-50)	74.3%	91.6%	24M		Bootstrap your own latent: A new approach to self-supervised Learning	deepmind/deepmind-research · alibaba/EasyCV · lucidrains/byol-pytorch · +28	2020
86	SimCLR (ResNet-50 2x)	74.2%	92.0%	94M		A Simple Framework for Contrastive Learning of Visual Representations	tensorflow/models · google-research/simclr · lightly-ai/lightly · +93	2020
86	Self-Classifier (ResNet-50)	74.2%	–	24M		Self-Supervised Classification Network	elad-amrani/self-classifier · abhsri/Reimplementation-of-Self-supervised-Classifier	2021
88	GroCo (ResNet-50)	73.9%	91.6	25M		Learning by Sorting: Self-supervised Learning with Group Ordering Constraints	ninatu/learning_by_sorting	2023
89	OBoW (ResNet-50)	73.8%	92.2%	24M		OBoW: Online Bag-of-Visual-Words Generation for Self-Supervised Learning	valeoai/obow · TRAILab/ST-SLidR · YufeiHU-fr/obow_ssl	2020
90	VICReg (ResNet50)	73.2	91.1	24M		VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learning	lightly-ai/lightly · vturrisi/solo-learn · facebookresearch/vicreg · +3	2021
90	Barlow Twins (ResNet-50)	73.2%	91	24M		Barlow Twins: Self-Supervised Learning via Redundancy Reduction	lightly-ai/lightly · facebookresearch/vissl · open-mmlab/mmselfsup · +21	2021
92	InfoMin (ResNet-50)	73.0%	91.1%	24M		What Makes for Good Views for Contrastive Learning?	HobbitLong/PyContrast	2020
93	DINO (ResMLP-24)	72.8%	–	30M		ResMLP: Feedforward networks for image classification with data-efficient training	rwightman/pytorch-image-models · xmu-xiaoma666/External-Attention-pytorch · facebookresearch/deit · +16	2021
93	MoBY (DeiT-S)	72.8%	–	22M		Self-Supervised Learning with Swin Transformers	microsoft/Swin-Transformer · alibaba/EasyCV · SwinTransformer/Transformer-SSL · +3	2021
95	I-VNE+ (ResNet-50)	72.1	91.0	25M		VNE: An Effective Method for Improving Deep Representation by Manipulating Eigenvalue Distribution	jaeill/CVPR23-VNE	2023
96	iGPT-XL (64x64, 15360 features)	72.0%	–	6801M		Generative Pretraining from Pixels	openai/image-gpt · EugenHotaj/pytorch-generative · teddykoker/image-gpt · +1	2020
97	SimCLRv2 (ResNet-50)	71.7%	90.4%	24M		Big Self-Supervised Models are Strong Semi-Supervised Learners	google-research/simclr · lightly-ai/lightly · sayakpaul/PAWS-TF · +6	2020
98	CPC v2 (ResNet-161) (arxiv v2)	71.5%	90.1%	305M		Data-Efficient Image Recognition with Contrastive Predictive Coding	philip-bachman/amdim-public · mf1024/Contrastive-Predictive-Coding-for-Image-Recognition-in-PyTorch · SeonghoBaek/FrameSequencePrediction · +1	2019
99	SimSiam (ResNet-50)	71.3%	–	24M		Exploring Simple Siamese Representation Learning	lightly-ai/lightly · open-mmlab/mmselfsup · vturrisi/solo-learn · +23	2020
100	MoCo v2 (ResNet-50)	71.1%	90.1%	24M		Improved Baselines with Momentum Contrastive Learning	open-mmlab/mmdetection · facebookresearch/moco · lightly-ai/lightly · +33	2020