Research map / Architectures and efficiency / Neural architecture search
Searching for MobileNetV3
Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, Hartwig Adam
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add h-swish instead of ReLU (ablation) | MobileNetV3-Large · ImageNet val | top-1 accuracy: 74.5 -> 75.2 | Improved |
| use optimized h-swish implementation (ablation) | MobileNetV3-Large · ImageNet val Pixel 1 | latency (ms): 57.5 -> 51.4 | Improved |
| enable h-swish from 16 channels layer instead of later (ablation) | MobileNetV3-Large · ImageNet val | top-1 accuracy: 75.2 -> 75.4 | Mixed |
| enable h-swish only from 112 channels layer (ablation) | MobileNetV3-Large · ImageNet val | top-1 accuracy: 75.2 -> 75.0 | Hurt |
| add LR-ASPP segmentation head instead of R-ASPP (ablation) | MobileNetV3 + RF2 · Cityscapes val | mIOU: 71.91 -> 72.37 | Improved |
| add LR-ASPP instead of R-ASPP (ablation) | MobileNetV2 + RF2 · Cityscapes val | mIOU: 72.56 -> 72.97 | Improved |
| reduce last-block channels by 2 (RF2) (ablation) | MobileNetV3 R-ASPP · Cityscapes val | mIOU: 72.64 -> 71.91 | Hurt |
| reduce segmentation head filters from 256 to 128 (ablation) | MobileNetV3 LR-ASPP · Cityscapes val | mIOU: 72.37 -> 72.36 | No clear effect |
| apply atrous convolution in last block (OS=16 vs 32) (ablation) | MobileNetV3-Large · Cityscapes test | mIOU: 72.0 -> 72.6 | Improved |
| reduce channels between C4 and C5 by 2 (ablation) | MobileNetV3-Large SSDLite · COCO test | mAP: 22.0 -> 22.0 | No clear effect |
| use MobileNetV3-Large instead of V2 | SSDLite · COCO test | mAP: 22.1 -> 22.0 | No clear effect |
| use MobileNetV3-Large instead of MobileNetV2 | ImageNet classification · ImageNet val | top-1 accuracy: 72.0 -> 75.2 | Improved |
Closest papers by meaning
- MobileNetV2: Inverted Residuals and Linear Bottlenecks (CVPR 2018 · similarity 0.89)
- MnasNet: Platform-Aware Neural Architecture Search for Mobile (CVPR 2019 · similarity 0.88)
- MnasFPN: Learning Latency-Aware Pyramid Architecture for Object Detection on Mobile Devices (CVPR 2020 · similarity 0.87)
- MobileNetV4: Universal Models for the Mobile Ecosystem (ECCV 2024 · similarity 0.86)
- MobileDets: Searching for Object Detection Architectures for Mobile Accelerators (CVPR 2021 · similarity 0.86)
- IFORMER: INTEGRATING CONVNET AND TRANSFORMER FOR MOBILE APPLICATION (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
