Image Classification Backbone Architectures (CNN and Vision Transformer)

Designing the core networks that turn an image into features, from residual CNNs to vision transformers.

Tier Grades

Tier A: Frontier AI labs

feasible

Big-tech labs introduced ViT and Swin with large pretraining sets.

Tier B: Indonesian AI labs

feasible

Training and comparing backbones on ImageNet-scale data is within an Indonesian lab reach.

Tier C: Institution units such as ITB

feasible

Many backbone papers come from small university teams; transfer to CIFAR-100 and VTAB fits a unit.

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

feasible

Pretrained backbones are the default starting point of any vision proof of concept.

Where this comes from

Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.

Anchors

ComputeViT needs substantially fewer computational resources to train than comparable CNNs when pretrained on large data, per the abstract; no figure is given there.
DataPretraining on large data, then transfer to ImageNet, CIFAR-100 and VTAB.
Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 12 authors on the ViT paper (arXiv 2010.11929).
Time horizonYears: classic backbones stay in use long after newer ones appear.
Shelf-life signalsA new architecture family (for example state-space or large-kernel) tops ImageNet, Backbone choice moves to pretrained foundation encoders
Pipeline roleTier A and academia publish backbones; Tier C benchmarks them on local data; Tier D picks pretrained ones as the base of client models.

Where this comes from

Team size is a proxy (author counts of exemplar papers), not a headcount.

Who and What

Exemplars

Vision Transformer (ViT) Meta

ConvNeXt

Swin Transformer ResNet50 (ImageNet)

Observed actors

Google

Meta

Microsoft AI

Linked models

+81

RepLKNetvisionproduced by

VAN (Visual Attention Network)visionproduced by

ConvMixervisionproduced by

Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences

UniFormervisionproduced by

ConvNeXtvisionproduced by

Swin S3 (AutoFormerV2)visionproduced by

MetaFormer (PoolFormer, ConvFormer, CAFormer)visionproduced by

Swin Transformer V2visionproduced by

HRFormervisionproduced by

Perceiver (vision)visionproduced by

VOLOvisionproduced by

Video Swin Transformervisionproduced by

ViT AugRegvisionproduced by

XCiTvisionproduced by

CoAtNetvisionproduced by NesTvisionproduced by

University of Chinese Academy of Sciences

Conformer (vision)visionproduced by

MLP-Mixervisionproduced by

Twinsvisionproduced by Visformervisionproduced by

MViT (Multiscale Vision Transformers)visionproduced by

CoaTvisionproduced by

CaiTvisionproduced by

PiT (Pooling-based ViT)visionproduced by

CvTvisionproduced by

ViViTvisionproduced by

CrossViTvisionproduced by

Swin-Bvisionproduced by

Swin-Tvisionproduced by

DPT (Dense Prediction Transformer)visionproduced by

ConViTvisionproduced by

ResNet-RSvisionproduced by

RegNetZvisionproduced by

TNT (Transformer in Transformer)visionproduced by

PVT (Pyramid Vision Transformer)visionproduced by

NFNetvisionproduced by

T2T-ViTvisionproduced by

Bottleneck Transformers (BoTNet)visionproduced by

RepVGGvisionproduced by

DeiTvisionproduced by

DeiT-B distilledvisionproduced by

DeiT-Tivisionproduced by

ViT-B/16visionproduced by

ViT-L/16visionproduced by MEAL V2visionproduced by

ResNeStvisionproduced by

RegNetvisionproduced by

TResNetvisionproduced by

BiT (Big Transfer)visionproduced by

CSPNetvisionproduced by

ECA-ResNetvisionproduced by HarDNetvisionproduced by

HRNetV2visionproduced by

ETRI (Electronics and Telecommunications Research Institute)

VoVNetvisionproduced by Res2Netvisionproduced by

Nanjing University of Science and Technology

Selective Kernel Networks (SK-ResNet)visionproduced by

University of Science and Technology of China (USTC)

HRNetvisionproduced by

IBN-Netvisionproduced by

ResNeXt-101 32x8d WSLvisionproduced by

SE-ResNet and SE-ResNeXt (Squeeze-and-Excitation)visionproduced by

DLA (Deep Layer Aggregation)visionproduced by

DPN (Dual Path Network)visionproduced by

ResNeXtvisionproduced by

ResNeXt-50 32x4dvisionproduced by

Xceptionvisionproduced by

DenseNetvisionproduced by

SimpleNet (classifier)visionproduced by Wide ResNetvisionproduced by

Inception v4visionproduced by

Inception-ResNet v2visionproduced by

Inception v3visionproduced by

Inception (GoogLeNet)visionproduced by

GoogLeNet (Inception v1)visionproduced by

VGGvisionproduced by

VGG-16visionproduced by

VGG-19visionproduced by

ZFNetvisionproduced by AlexNetvisionproduced by CaffeNetvisionproduced by

InternImagevisionproduced by

MUSIQvisionproduced by

Links

R&D TopicImage Classification Backbone Architectures (CNN and Vision Transformer)

ModelsResNet50 (ImageNet)vision

ConvNeXtMeta, vision

Vision Transformer (ViT)Google, vision

Swin TransformerMicrosoft AI, vision+101 more

Use Casesnone recorded

Who Runs Itnone recorded

Graph

Models 105

HRNetUniversity of Science and Technology of China (USTC), Microsoft Research, visionproduced by

Inception (GoogLeNet)Google, visionproduced by

DeiTMeta, visionproduced by

DINOv3 ConvNeXt-BMeta, visionproduced by

DINOv3 ConvNeXt-LMeta, visionproduced by

DINOv3 ConvNeXt-SMeta, visionproduced by

DINOv3 ConvNeXt-TMeta, visionproduced by

HRFormerPeking University, Microsoft Research, visionproduced by

ResNeXtUC San Diego, Meta, visionproduced by

Swin TransformerMicrosoft AI, visionexemplar

Vision Transformer (ViT)Google, visionexemplar

ConvNeXtMeta, visionproduced by

+93

ConvNeXtMeta, visionexemplar

ConvNeXt V2Meta, visionproduced by

DeiT IIIMeta, visionproduced by

DeiT-B distilledMeta, visionproduced by

DenseNetCornell University, Tsinghua University, Meta, visionproduced by

HRNetV2Microsoft Research, visionproduced by

ResNeXt-50 32x4dUC San Diego, Meta, visionproduced by

Swin Transformer V2Microsoft AI, visionproduced by

Swin-BMicrosoft AI, visionproduced by

UniFormerShenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shanghai AI Lab, visionproduced by

VGGUniversity of Oxford, visionproduced by

Bottleneck Transformers (BoTNet)UC Berkeley, Google, visionproduced by

Conformer (vision)University of Chinese Academy of Sciences, Peng Cheng Laboratory, visionproduced by

ConViTEcole Normale Superieure (Paris), Meta, visionproduced by

ConvMixerCarnegie Mellon University, Bosch, visionproduced by

DaViTThe University of Hong Kong, Microsoft, visionproduced by

DeiT-TiMeta, visionproduced by

DPT (Dense Prediction Transformer)Intel, visionproduced by

GoogLeNet (Inception v1)Google, visionproduced by

HorNetTsinghua University, Meta, visionproduced by

IBN-NetThe Chinese University of Hong Kong, SenseTime, visionproduced by

Inception v3Google, visionproduced by

Inception v4Google, visionproduced by

Inception-ResNet v2Google, visionproduced by

InceptionNeXtNational University of Singapore, Sea AI Lab, visionproduced by

InternImageShanghai AI Lab, Tsinghua University, visionproduced by

MaxViTGoogle, University of Texas at Austin, visionproduced by

MetaFormer (PoolFormer, ConvFormer, CAFormer)Sea AI Lab, National University of Singapore, visionproduced by

MViT (Multiscale Vision Transformers)Meta, visionproduced by

RegNetMeta, visionproduced by

RegNetZMeta, visionproduced by

RepLKNetTsinghua University, Megvii, visionproduced by

RepVGGTsinghua University, Megvii, visionproduced by

ResNeStMeta, Amazon, visionproduced by ResNet50 (ImageNet)visionexemplar

ResNeXt-101 32x8d WSLMeta, visionproduced by

RIFormerShanghai AI Lab, The University of Hong Kong, visionproduced by

SE-ResNet and SE-ResNeXt (Squeeze-and-Excitation)Momenta, University of Oxford, visionproduced by

Selective Kernel Networks (SK-ResNet)Nanjing University of Science and Technology, Momenta, visionproduced by

Swin S3 (AutoFormerV2)Stony Brook University, Microsoft Research, visionproduced by

Swin-TMicrosoft AI, visionproduced by

TwinsMeituan, University of Adelaide, visionproduced by

VGG-16University of Oxford, visionproduced by

VGG-19University of Oxford, visionproduced by

Video Swin TransformerMicrosoft Research, visionproduced by

ViT AugRegGoogle, visionproduced by

ViT-B/16Google, visionproduced by

ViT-L/16Google, visionproduced by

ViViTGoogle, visionproduced by

VOLOSea AI Lab, National University of Singapore, visionproduced by

BiT (Big Transfer)Google, visionproduced by

CaiTMeta, visionproduced by

CoaTUC San Diego, visionproduced by

CoAtNetGoogle, visionproduced by

CrossViTIBM, visionproduced by

CSPNetAcademia Sinica, visionproduced by

CvTMicrosoft AI, visionproduced by

DLA (Deep Layer Aggregation)UC Berkeley, visionproduced by

DPN (Dual Path Network)National University of Singapore, visionproduced by

ECA-ResNetTianjin University, visionproduced by

FocalNetMicrosoft Research, visionproduced by

GCViTNVIDIA, visionproduced by Hieravisionproduced by

MambaOutNational University of Singapore, visionproduced by

MLP-MixerGoogle, visionproduced by

MUSIQGoogle, visionproduced by

NaViT and FlexiViT (NaFlexViT)Google DeepMind, visionproduced by

NFNetGoogle DeepMind, visionproduced by

Perceiver (vision)Google DeepMind, visionproduced by

PiT (Pooling-based ViT)Naver, visionproduced by

PVT (Pyramid Vision Transformer)Nanjing University, visionproduced by

RDNet (DenseNets Reloaded)Naver, visionproduced by

ResNet-RSGoogle, visionproduced by

RoPE-ViTNaver, visionproduced by

Sequencer2DRikkyo University, visionproduced by

SimpleNet (classifier)Islamic Azad University, visionproduced by

SNN-MLPHuawei (Huawei Cloud), visionproduced by

T2T-ViTNational University of Singapore, visionproduced by

TNT (Transformer in Transformer)Huawei, visionproduced by

TResNetAlibaba DAMO, visionproduced by

VAN (Visual Attention Network)Tsinghua University, visionproduced by

VoVNetETRI (Electronics and Telecommunications Research Institute), visionproduced by

XceptionGoogle, visionproduced by

XCiTMeta, visionproduced by

ZFNetNew York University, visionproduced by AlexNetvisionproduced by CaffeNetvisionproduced by HarDNetvisionproduced by MEAL V2visionproduced by NesTvisionproduced by Res2Netvisionproduced by Visformervisionproduced by Wide ResNetvisionproduced by

Papers

Deep Residual Learning for Image RecognitionKaiming He et al., arXiv preprint, 2015 founding
An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy et al., arXiv preprint, 2020 founding
A ConvNet for the 2020sZhuang Liu et al., CVPR 2022 (per arXiv comment), 2022 state of the art
Swin Transformer: Hierarchical Vision Transformer using Shifted WindowsZe Liu et al., arXiv preprint, 2021 state of the art

Where this comes from

The papers to read first: where the direction began, its current best, and a survey when one exists.

Sources

arxiv.org/abs/2010.11929 https://arxiv.org/abs/2010.11929

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Image Classification Backbone Architectures (CNN and Vision Transformer)", https://aurion.semai.id/topics/vision-backbone-architectures/, records of October 2026.