DSPC-Net: A Dual-Branch Self-Correcting Pose Consistency Network for Head Pose Estimation
Abstract
Existing head pose estimation (HPE) methods focus on utilizing the information of one single image and ignore cross-viewpoint pose consistency. To remedy this, we propose a dual-branch self-correcting pose consistency network (DSPC-Net) for HPE. The idea is to construct flipped image pairs as intrinsic pose constraints and leverage a weight-shared dual-branch network via self-correcting learning for precise pose estimation. To this end, we design a multi-level hybrid feature interaction fusion module to integrate local and global features for improved robustness. We build a learnable auxiliary matrix to calibrate and fuse the predicted rotation matrices by considering their geometric consistency. To our best knowledge, we are the first to use such self-correcting learning for HPE. We also introduce a multi-loss function to overcome the limitations of conventional loss functions. Experiments on two datasets demonstrate the advantages of DSPC-Net over state-of-the-art methods for HPE. Specifically, DSPC-Net achieves an MAE (mean absolute error) of Euler angles of 4.00 degrees and an MAEV (MAE of vectors) of 4.96 on the AFLW2000 dataset, with an MAE of 3.65 degrees and an MAEV of 5.02 on the BIWI dataset.