Pose-Invariant Face Recognition via Adaptive Angular Distillation

Abstract

Pose-invariant face recognition is a practically useful but challenging task. This paper introduces a novel method to learn pose-invariant feature representation without normalizing profile faces to frontal ones or learning disentangled features. We first design a novel strategy to learn pose-invariant feature embeddings by distilling the angular knowledge of frontal faces extracted by teacher network to student network, which enables the handling of faces with large pose variations. In this way, the features of faces across variant poses can cluster compactly for the same person to create a pose-invariant face representation. Secondly, we propose a Pose-Adaptive Angular Distillation loss to mitigate the negative effect of uneven distribution of face poses in the training dataset to pay more attention to the samples with large pose variations. Extensive experiments on two challenging benchmarks (IJB-A and CFP-FP) show that our approach consistently outperforms the existing methods.

MaskLRF: Self-Supervised Pretraining via Masked Autoencoding of Local Reference Frames for Rotation-Invariant 3D Point Set Analysis

“…Various RI DNNs for 3D point sets have been proposed. They can be classified into the following three approaches; extracting inherently RI feature ( [31], [32], [33], [34], [35], [36], [37]), designing rotation-equivariant DNN architecture ( [38], [39]), and normalizing rotation of 3D point sets ( [40], [41], [42], [43], [44], [45], [46], [47]).…”

“…Normalizing rotation. The studies in this category achieve rotation invariance by normalizing the orientation of a 3D point set at global scale [37], [42] or local scale [40], [43], [47]. Compared to global scale, rotation normalization at local scale is easier since 3D shape in a local region tends to be simple.…”

“…Specifically, three mutually orthogonal axes for rotation normalization are computed by eigendecomposition of the covariance matrix of the 3D points in a local region. Luo et al [43] and Zhang et al [47] propose a DNN block that predicts an intrinsic LRF of a local region for its rotation-normalization. The use of LRF achieves rotational invariance without loss of 3D shape information in a local region.…”

“…The token features output from the last encoding Transformer block are upsampled to pointwise features. This paper adopts Poseaware Feature Propagation devised in [47] for upsampling. This module leverages not only distances among neighboring 3D points but also angles among LRFs.…”

“…They are six MPM-based methods [6], [7], [8], [9], [10], [11]. In addition, previously proposed DNNs having rotation invariance [35], [36] 1 , [37], [39], [40], [42], [43], [45], [47], [60] are included in the set of competitors. These RI DNNs are not pretrained, but are trained from scratch for each downstream task.…”