Abstract | ||
---|---|---|
3D face recognition has developed rapidly in the past decade for its robustness to large variations of illumination and pose. Plenty of researches in the past have been done based on high-quality 3D faces, which, however, have limitations in real applications because of expensive 3D scanners and high computational cost. Recently, low-quality 3D face recognition is attracting increasing attention because of its lower acquisition cost and faster acquisition speed, making it easier to generalize in practice. However, identity feature in low-quality 3D face data is easily damaged by massive noise, leading to very low accuracy in face recognition. To deal with this problem, we propose a trainable Soft Thresholding Module (STM) to explicitly recover 3D faces from highly noised inputs, which is different from existing methods that either use untrainable preprocessing techniques or implicitly learn robust feature representations. Experimental results show that the proposed method is effective and achieves state-of-the-art performance in low-quality 3D face recognition. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1007/978-3-030-86608-2_46 | BIOMETRIC RECOGNITION (CCBR 2021) |
Keywords | DocType | Volume |
3D face recognition, Soft thresholding module, Low-quality | Conference | 12878 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Shudi Xiao | 1 | 0 | 0.34 |
Shuiwang Li | 2 | 0 | 0.34 |
Qijun Zhao | 3 | 419 | 38.37 |