RGBD-based Hardware Friendly Head Pose Estimation System via Convolutional attention module

Abstract

Head pose estimation from RGB images without depth information is a challenging task, owing to the loss of spatial information and large head pose variations in the wild. However, most studies adopt deeper convolutional neural network (CNN) models, such as ResNet50, which are limited by the enormous number of parameters to be implemented on edge devices. Owing to novel technological advancements, several edge devices have also included depth cameras and obtain high-quality images. In this study, we propose a lightweight CNN for head pose estimation. By adopting attention module and feature decoupler, we resume the performance decreasing by lower parameters. Moreover, we classify the ground-truth head pose angles of the model intermittently, and adopt the multi-loss strategy to train our model. We evaluate the proposed method on three challenging benchmark datasets, and achieved optimal results for Yaw pose and average. The obtained results indicate that although the proposed model has less parameters, it still maintains a remarkable performance. The total number of parameters is 0.19 M, including RGB and depth path, which is 50% lower than FSA-Net. Consequently, the inference speed is 0.92 ms per pair RGB-D images, which is 8% faster than FSA-Net. With fewer parameters, we achieved 3.1 MAE on yaw angle, which is 22.69% lower than that of Quatnet, including 3.5 MAE on average, which is 7.40% lower than those of other advanced methods.