Head pose estimation in the wild using Convolutional Neural Networks and adaptive gradient methods

Abstract

Head pose estimation is an old problem that is recently receiving new attention because of possible applications in human-robot interaction, augmented reality and driving assistance. However, most of the existing work has been tested in controlled environments and is not robust enough for real-world applications. In order to handle these limitations we propose an approach based on Convolutional Neural Networks (CNNs) supplemented with the most recent techniques adopted from the deep learning community. We evaluate the performance of four architectures on recently released in-the-wild datasets. Moreover, we investigate the use of dropout and adaptive gradient methods giving a contribution to their ongoing validation. The results show that joining CNNs and adaptive gradient methods leads to the state-of-the-art in unconstrained head pose estimation.

Towards Real-Time Head Pose Estimation: Exploring Parameter-Reduced Residual Networks on In-the-wild Datasets

“…In comparison, our ResNet18-64 has the lowest number of parameters while predicting more accurately than the LeNet-5 variant [25] and nearly as accurate as the ResNet50 [27]. Patacchiola and Cangelosi [25] also use low-resolution images with 64 x 64 pixels, while Ruiz et al [27] take larger images with 224 x 224 pixels. To improve the computational efficiency, we believe that low-resolution images are better suited for real-world applications.…”

“…The number of parameters of [27] is based on their provided open source implementation, which is executable on a GPU based system. 4 In order to compare the frame rate, we reimplemented the LeNet-5 variant of [25]. In comparison, our ResNet18-64 has the lowest number of parameters while predicting more accurately than the LeNet-5 variant [25] and nearly as accurate as the ResNet50 [27].…”

“…4 In order to compare the frame rate, we reimplemented the LeNet-5 variant of [25]. In comparison, our ResNet18-64 has the lowest number of parameters while predicting more accurately than the LeNet-5 variant [25] and nearly as accurate as the ResNet50 [27]. Patacchiola and Cangelosi [25] also use low-resolution images with 64 x 64 pixels, while Ruiz et al [27] take larger images with 224 x 224 pixels.…”

“…In contrast to MLPs, CNNs display a high tolerance to shift and distortion variance. There are several recent approaches that employ the in-the-wild datasets AFLW and AFW for training and use the AFW dataset as a testing benchmark: Patacchiola and Cangelosi [25] compare various LeNet-5 [19] variants trained with different gradient and adaptive gradient methods. Ruiz et al [27] train a 50-layer ResNet with a combined loss function of mean squared loss and cross entropy loss for all three angles.…”

“…Furthermore, few images are provided for extreme angles. Following the approach of Patacchiola and Cangelosi [25], we filter the dataset and only keep images in the following label ranges: ±100 • for the yaw angle, ±45 • for the pitch angle and ±25 • for the roll angle. The yaw angle of the AFW dataset is also restricted to ±100 • , as this angle is used for testing the trained networks.…”