PointFace: Point Cloud Encoder-Based Feature Embedding for 3-D Face Recognition

Point CNN:3D Face Recognition with Local Feature Descriptor and Feature Enhancement Mechanism

“…The results are shown in Table 9. According to the results, in the NU subset, which has no other interference, Jiang et al [31] achieves best accuracy, but in the OC and PS subsets, our method achieves the best accuracy, which proves that our network is better to cope with partial occlusions and head pose interference. Figure 14 shows the t-SN example of our network for face recognition on three datasets (each dataset selects five subjects for classification and each color represents one subject).…”

“…We divide the training set and test set of Lock3DFace according to the method in [ 31 ], in which the 340 subjects are randomly selected as the training set and the remaining 169 subjects are selected as the test set.…”

“…The results of ablation experiments prove the effectiveness of our proposed method. In this section, according to [ 31 ], we use our best results to conduct comparison experiments with current advanced methods on three public datasets and analyze the results.…”

“…Dutta et al [23] 7 M.S et al [38] 6 Xiao et al [1] 117 Xiao et al [32] 125 Mu et al [22] 136 Jiang et al [31] 118 Ours * 139 Ours 125…”

“…Yu et al [ 30 ] proposed a meta learning-based adversarial training (MLAT) algorithm for deep 3D face recognition on point clouds, which consists of two alternate modules: adversarial sample generating for 3D face data augmentation and meta learning-based deep network training. Jiang et al [ 31 ] used two weight-shared encoders and a feature similarity loss to guide the encoders to obtain discriminative face representations and have achieved good performance on different datasets. Apart from face recognition, point clouds are also used for other 3D face analysis tasks such as face verification and head pose estimation [ 1 , 2 , 32 ].…”