Reproduce “MobileDets: Searching for Object Detection Architectures for Mobile Accelerators” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunyang Xiong, Hanxiao Liu, Suyog Gupta, Berkin Akin, Gabriel Bender, Yongzhe Wang, Pieter-Jan Kindermans, Mingxing Tan and 2 more
Researchers from Google and the University of Wisconsin-Madison developed MobileDets, a family of object detection architectures optimized for various mobile hardware accelerators through neural architecture search. These models incorporate a broader set of convolutional operations beyond traditional inverted bottlenecks, achieving improved accuracy-latency trade-offs on mobile CPUs, EdgeTPUs, DSPs, and edge GPUs.
Key insights
- Regular (full) convolutions, traditionally considered less efficient, can execute significantly faster than depthwise convolutions on specialized mobile accelerators like EdgeTPUs and DSPs for certain configurations.
- Expanding the neural architecture search space beyond inverted bottleneck layers by including fused and Tucker convolutions is crucial for discovering architectures that yield superior accuracy-latency trade-offs on mobile accelerators.
- Optimal network architectures vary substantially across different hardware platforms, with EdgeTPU and DSP-optimized MobileDets extensively incorporating regular convolutions, indicating the necessity for hardware-specific design.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.