Reproduce “ELSR: Extreme Low-Power Super Resolution Network For Mobile Devices” 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
Tianyu Xu, Zhuang Jia, Yijian Zhang, Long Bao, Heng Sun
An Extreme Low-Power Super Resolution (ELSR) network was developed to enable real-time video super-resolution on mobile devices with minimal energy usage. This model secured first place in the Mobile AI & AIM 2022 challenge by achieving a PSNR of 27.34 dB while consuming only 0.09 W/30FPS on a MediaTek Dimensity 9000 platform.
Key insights
- Simple, plain convolutional architectures demonstrate superior energy efficiency compared to complex multi-branch or attention-based designs for mobile VSR, despite minor PSNR differences.
- A carefully designed multi-stage training strategy, including adaptive weight transfer, is crucial for accelerating convergence and achieving optimal performance in extremely compact networks.
- Direct optimization for actual power consumption on a target mobile hardware platform (e.g., MediaTek Dimensity 9000) leads to different architectural design choices compared to optimizing for FLOPs or desktop GPU speed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.