کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / سیزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

A Comparison between Slimed Network and Pruned Network for Head Pose Estimation

نویسندگان :

Amir Salimiparsa¹ Hadi Veisi² Mohammad-shahram Moin³

1- دانشگاه تهران 2- دانشگاه تهران ٫ دانشکده علوم و فنون نوین 3- پژوهشگاه ارتباطات وفناوری اطلاعات

کلمات کلیدی :

Head pose estimation،MobileNet،Pruning،Quantization،Deep neural networks

چکیده :

Head pose estimation is a critical problem with a wide range of applications. There are many methods that almost solved head pose estimation problems but they are computationally expensive and not suitable for edge devices and embedded systems. In this paper, a deep learning network based on a modified MobileNetV3 architecture is proposed to reduce the computational cost with results comparable to heavy methods. The proposed method is pruned to achieve even less computational cost and results in a network that is more ideal for edge devices and smartphones. The architecture used is MobileNetV3Small which has more inverted residual blocks, making it able to inherit MobileNetV3Large performance but with less width, followed by dense layers. Pruning is enhanced by estimating layer importance and resource reallocation, in order for the informative layers to be less affected by pruning and also to improve performance. In the experiments, the proposed model performs better than many existing heavies with 3.46 MAE before the pruning and 3.61 MAE after the pruning, even though the model has six times fewer parameters than the others and its inference time is about 7ms.