Research map / Detection and segmentation / CNN features
MobileNetV2: Inverted Residuals and Linear Bottlenecks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use MobileNetV2 instead of MobileNetV1 | ImageNet classification · ImageNet | Top-1 accuracy: 70.6 -> 72.0 | Helped |
| Use MobileNetV2 instead of MobileNetV1 | ImageNet inference on Pixel 1 · ImageNet | CPU latency: 113ms -> 75ms | Helped |
| Replace regular SSD prediction-layer convolutions with depthwise separable convolutions (ablation) | SSD detection head · COCO detection | MAdds: 1.25B -> 0.35B | Helped |
| Replace regular SSD prediction-layer convolutions with depthwise separable convolutions (ablation) | SSD detection head · COCO detection | Parameters: 14.8M -> 2.1M | Helped |
| Use MobileNetV2 instead of MobileNetV1 as the SSDLite feature extractor | SSDLite · COCO test-dev | MAdds: 1.3B -> 0.8B | Helped |
| Use MobileNetV2 instead of MobileNetV1 as the SSDLite feature extractor | SSDLite · COCO test-dev | mAP: 22.2 -> 22.1 | Mixed |
| Use linear rather than nonlinear narrow bottleneck layers (ablation) | MobileNetV2 · Image classification | classification performance | Helped |
| Connect residual shortcuts between bottlenecks rather than expanded layers (ablation) | MobileNetV2 · Image classification | classification performance | Helped |
| Vary the expansion factor within 5–10 (ablation) | MobileNetV2 · Image classification | classification performance | No clear effect |
| Remove ASPP from the OS 16 DeepLabv3 head (ablation) | MobileNetV2* + DeepLabv3 · PASCAL VOC 2012 validation | mIOU: 75.70 -> 75.32 | Hurt |
| Remove ASPP from the OS 16 DeepLabv3 head (ablation) | MobileNetV2* + DeepLabv3 · PASCAL VOC 2012 validation | MAdds: 5.8B -> 2.75B | Helped |
| Use OS 8 with multi-scale and flipped test inputs rather than OS 16 single-scale inference (ablation) | MobileNetV2* + DeepLabv3 with ASPP · PASCAL VOC 2012 validation | mIOU: 75.70 -> 78.42 | Helped |
Closest papers by meaning
- MobileDets: Searching for Object Detection Architectures for Mobile Accelerators (CVPR 2021 · similarity 0.88)
- Mobile-Former: Bridging MobileNet and Transformer (CVPR 2022 · similarity 0.88)
- MOAT: Alternating Mobile Convolution and Attention Brings Strong Vision Models (ICLR 2023 · similarity 0.88)
- IFORMER: INTEGRATING CONVNET AND TRANSFORMER FOR MOBILE APPLICATION (ICLR 2025 · similarity 0.86)
- MobileInst: Video Instance Segmentation on the Mobile (AAAI 2024 · similarity 0.84)
- Recognize Any Regions (NeurIPS 2024 · similarity 0.84)
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.