Efficient Vision Networks for Mobile and Edge Devices

Small, fast vision networks, often found by architecture search, that run on phones, cameras and microcontrollers.

Tier Grades

Tier A: Frontier AI labs

feasible

Architecture search at Google scale produced EfficientNet and MobileNet.

Tier B: Indonesian AI labs

feasible

Training small networks on ImageNet is feasible for an Indonesian lab.

Tier C: Institution units such as ITB

feasible

A 2-author paper set this line, so a university unit can contribute.

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

feasible

Small models that run on site cameras and phones suit plantation and factory pilots with limited connectivity.

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

ComputeEfficientNet-B7 is 8.4x smaller and 6.1x faster at inference than the best earlier ConvNet; training compute is not stated in the abstract.
DataImageNet (84.3% top-1 for B7), with transfer to CIFAR-100 (91.7%) and Flowers (98.8%).
Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 2 authors on the EfficientNet paper (arXiv 1905.11946).
Time horizonMonths to a year.
Shelf-life signalsNew edge accelerators favour different operators, Quantisation-aware designs beat larger float models on device
Pipeline roleTier A searches and releases efficient networks; Tier C ports them to local hardware; Tier D deploys them on edge devices in pilots.

Where this comes from

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

Who and What

Exemplars

EfficientNet Google

MobileNet v4 Megvii

ShuffleNet v2 Apple

Observed actors

Linked models

+63

PP-LCNetvisionproduced by

ViPNASvisionproduced by

SCRFDvisionproduced by

Huazhong University of Science and Technology

Lite-HRNetvisionproduced by

LeViTvisionproduced by

EfficientNetV2visionproduced by

MoViNetvisionproduced by

HardCoRe-NASvisionproduced by

Objectron 3D object detectionvisionproduced by

ReXNetvisionproduced by

GPU-Efficient Networks (GENet)visionproduced by

MediaPipe Handsvisionproduced by

BlazePose GHUMvisionproduced by

X3Dvisionproduced by

SINet (portrait segmentation)visionproduced by

GhostNetvisionproduced by

EfficientDet-Litevisionproduced by

Once-for-All (OFA)visionproduced by MixNetvisionproduced by

Face Meshvisionproduced by

BlazeFacevisionproduced by

SelecSLSvisionproduced by

EfficientNet-B0visionproduced by

EfficientNet-B7visionproduced by

EfficientNet-Litevisionproduced by

MobileNet v3visionproduced by

OSNetvisionproduced by

Single-Path NASvisionproduced by

ProxylessNASvisionproduced by

MnasNetvisionproduced by MobileFaceNetvisionproduced by

SSD MobileNet v2visionproduced by

PNASNet-5visionproduced by

NASNetvisionproduced by

ShuffleNetvisionproduced by

MobileNetvisionproduced by

MobileNet v1visionproduced by

SqueezeNetvisionproduced by

MobileNet v5visionproduced by

MobileNet-EdgeTPUvisionproduced by

AmoebaNet-Avisionproduced by PP-HGNetvisionproduced by

AutoSeg-EdgeTPUvisionproduced by

FastDepthvisionproduced by

MoveNetvisionproduced by

PoseNetvisionproduced by

MediaPipe vision modelsvisionproduced by

MediaPipe Selfie Segmentervisionproduced by Ultra-Light-Fast-Generic-Face-Detector (UltraFace)visionproduced by YuNetvisionproduced by

DamoFDvisionproduced by

Tiny Video Netvisionproduced by

QuickSRNetvisionproduced by XLSR (super-resolution)visionproduced by

Intel binary, sparse and retrieval modelsvisionproduced by

Qualcomm face modelsvisionproduced by

AIY food, product and nature classifiersvisionproduced by

Mobile Object Localizervisionproduced by

HBONetvisionproduced by

YOLOlitevisionproduced by

YOLO-Provisionproduced by

XiNet Picasso Muse style transferimage generationproduced by Fast-SRGANimage generationproduced by

Links

R&D TopicEfficient Vision Networks for Mobile and Edge Devices

Models

MobileNet v4Google, vision

EfficientNetGoogle, vision

ShuffleNet v2Megvii, vision

MobileOneApple, vision+83 more

Use Casesnone recorded

Who Runs Itnone recorded

Graph

Models 87

MediaPipe vision modelsGoogle, visionproduced by

MobileNetGoogle, visionproduced by

EfficientNetGoogle, visionexemplar

AutoSeg-EdgeTPUGoogle, visionproduced by

EfficientDet-LiteGoogle, visionproduced by

Lite-HRNetHuazhong University of Science and Technology, Microsoft Research, visionproduced by

MobileNet v3Google, visionproduced by

BlazeFaceGoogle, visionproduced by

BlazePose GHUMGoogle, visionproduced by

EdgeTAMMeta, visionproduced by

EfficientNet-B0Google, visionproduced by

EfficientNet-LiteGoogle, visionproduced by

+75

EfficientNetV2Google, visionproduced by

Face MeshGoogle, visionproduced by

MediaPipe HandsGoogle, visionproduced by

MediaPipe Selfie SegmenterGoogle, visionproduced by

MobileNet v1Google, visionproduced by

MobileNet v4Google, visionexemplar

MobileNet v5Google, visionproduced by

MobileNet-EdgeTPUGoogle, visionproduced by

OSNetUniversity of Surrey, Queen Mary University of London, visionproduced by

PP-OCRv5 mobileBaidu, OCR and documentproduced by

SSD MobileNet v2Google, visionproduced by

ViPNASThe Chinese University of Hong Kong, SenseTime, visionproduced by

DamoFDAlibaba DAMO, visionproduced by

EfficientFormerSnap Inc., Northeastern University, visionproduced by

EfficientNet-B7Google, visionproduced by

EfficientViT (cascaded group attention)The Chinese University of Hong Kong, Microsoft Research, visionproduced by

EfficientViT-SAMMIT Han Lab, visionproduced by

FastDepthMassachusetts Institute of Technology, visionproduced by

Intel binary, sparse and retrieval modelsIntel, visionproduced by

Mobile Object LocalizerGoogle, visionproduced by

MobileSAMKyung Hee University, visionproduced by

MoveNetGoogle, visionproduced by

MoViNetGoogle, visionproduced by

Objectron 3D object detectionGoogle, visionproduced by

PoseNetGoogle, visionproduced by

Qualcomm face modelsQualcomm, visionproduced by

QuickSRNetQualcomm, visionproduced by

SCRFDInsightFace (Deep Insight), visionproduced by

ShuffleNetMegvii, visionproduced by

ShuffleNet v2Megvii, visionexemplar

SINet (portrait segmentation)Naver, visionproduced by

SlimSAMNational University of Singapore, visionproduced by

Tiny Video NetGoogle, visionproduced by

X3DMeta, visionproduced by

XiNet Picasso Muse style transferSTMicroelectronics, image generationproduced by

YOLO-ProEdge Impulse, visionproduced by

YOLOliteRoboflow, visionproduced by

AIY food, product and nature classifiersGoogle, visionproduced by

AmoebaNet-AGoogle, visionproduced by

CPUBoneUniversity of Udine, visionproduced by

EdgeNeXtMBZUAI, visionproduced by

EfficientViT (multi-scale linear attention)Massachusetts Institute of Technology, visionproduced by Fast-SRGANimage generationproduced by

FasterNetHong Kong University of Science and Technology, visionproduced by

FastViTApple, visionproduced by

GhostNetHuawei (Huawei Cloud), visionproduced by

GPU-Efficient Networks (GENet)Alibaba Group, visionproduced by

GPUNetNVIDIA, visionproduced by

HardCoRe-NASAlibaba Group, visionproduced by

HBONetIntel, visionproduced by

LeViTMeta, visionproduced by

MnasNetGoogle, visionproduced by MobileFaceNetvisionproduced by

MobileOneApple, visionexemplar

MobileViTApple, visionproduced by

NASNetGoogle, visionproduced by

Next-ViTByteDance, visionproduced by

Once-for-All (OFA)MIT Han Lab, visionproduced by

PNASNet-5Google, visionproduced by

PP-LCNetBaidu, visionproduced by

ProxylessNASMIT Han Lab, visionproduced by

RepViTTsinghua University, visionproduced by

ReXNetNaver, visionproduced by

SelecSLSMax Planck Institute for Informatics, visionproduced by

SHViTUniversity of Seoul, visionproduced by

Single-Path NASCarnegie Mellon University, visionproduced by

SqueezeNetUC Berkeley, visionproduced by

StarNetNortheastern University, visionproduced by

SwiftFormerMBZUAI, visionproduced by

TinyViTMicrosoft Research, visionproduced by Ultra-Light-Fast-Generic-Face-Detector (UltraFace)visionproduced by XLSR (super-resolution)visionproduced by YuNetvisionproduced by MixNetvisionproduced by PP-HGNetvisionproduced by

Papers

MobileNetV2: Inverted Residuals and Linear BottlenecksMark Sandler et al., The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp., 2018 founding
EfficientNet: Rethinking Model Scaling for Convolutional Neural NetworksMingxing Tan et al., International Conference on Machine Learning, 2019, 2019 founding
MobileNetV4 -- Universal Models for the Mobile EcosystemDanfeng Qin et al., arXiv preprint, 2024 state of the art
EfficientViT: Multi-Scale Linear Attention for High-Resolution Dense PredictionHan Cai et al., ICCV 2023 (per arXiv comment), 2022 state of the art
Searching for MobileNetV3Andrew Howard et al., ICCV 2019 (per arXiv comment), 2019 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/1905.11946 https://arxiv.org/abs/1905.11946

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Efficient Vision Networks for Mobile and Edge Devices", https://aurion.semai.id/topics/efficient-vision-networks-for-edge/, records of October 2026.