Optimising 3D-CNN Design towards Human Pose Estimation on Low Power Devices

Manolis Vasileiadis (Imperial College London), Christos-Savvas Bouganis (Imperial College London), Georgios Stavropoulos (Centre for Research and Technology, Hellas, Information Technologies Institute), Dimitrios Tzovaras (Centre for Research and Technology, Hellas)

3D CNN-based architectures have found application in a variety of 3D vision tasks, significantly outperforming earlier approaches. This increase in accuracy, however, has come at the cost of computational complexity, with deep learning models becoming more and more complex, requiring significant computational resources, especially in the case of 3D data. Meanwhile, the growing adoption of low power devices in various technology fields has shifted the research focus towards the implementation of deep learning on systems with limited resources. While plenty of approaches have achieved promising results in terms of reducing the computational complexity in 2D tasks, their applicability in 3D-CNN designs has not been thoroughly researched. The current work aims at filling this void, by investigating a series of efficient CNN design techniques within the scope of 3D-CNNs, in order to produce guidelines for 3D-CNN design that can be applied to already established architectures, reducing their computational complexity. Following these guidelines, a computationally efficient 3D-CNN architecture for human pose estimation from 3D data is proposed, achieving comparable accuracy to the state-of-the-art. The proposed design guidelines are further validated within the scope of 3D object classification, achieving high accuracy results at a low computational cost.


Paper (PDF)
Supplementary material (ZIP)

title={Optimising 3D-CNN Design towards Human Pose Estimation on Low Power Devices},
author={Manolis Vasileiadis and Christos-Savvas Bouganis and Georgios Stavropoulos and Dimitrios Tzovaras},
booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
publisher={BMVA Press},
editor={Kirill Sidorov and Yulia Hicks},