Towards Efficient U-Nets: A Coupled and Quantized Approach

Zhiqiang Tang, Xi Peng, Kang Li, Dimitris N. Metaxas

Abstract: In this paper, we propose to couple stacked U-Nets for efficient visual landmark localization. The key idea is to globally reuse features of the same semantic meanings across the stacked U-Nets. The feature reuse makes each U-Net light-weighted. Specially, we propose an <inline-formula><tex-math notation="LaTeX">$order$</tex-math></inline-formula> - <inline-formula><tex-math notation="LaTeX">$K$</tex-math></inline-formula> coupling design to trim off long-distance shortcuts, together with an iterative refinement and memory sharing mechanism. To further improve the efficiency, we quantize the parameters, intermediate features, and gradients of the coupled U-Nets to low bit-width numbers. We validate our approach in two tasks: human pose estimation and facial landmark localization. The results show that our approach achieves state-of-the-art localization accuracy but using <inline-formula><tex-math notation="LaTeX">$\sim 70\%$</tex-math></inline-formula> fewer parameters, <inline-formula><tex-math notation="LaTeX">$\sim 30\%$</tex-math></inline-formula> less inference time, <inline-formula><tex-math notation="LaTeX">$\sim 98\%$</tex-math></inline-formula> less model size, and saving <inline-formula><tex-math notation="LaTeX">$\sim 75\%$</tex-math></inline-formula> training memory compared with benchmark localizers.