Multi-Attribute Dynamic Learning for 3D Face Reconstruction and Dense Alignment in the Unconstrained Environment

3D face reconstruction and dense alignment have been challenging problems in unconstrained environments. Existing methods usually designed neural networks to regress the 62 facial attribute parameters that determine identity, expression, and pose, while the different sensitivities of different facial attributes to the varying unconstrained environment were not considered. To exploit this for the purpose of improving the dense alignment and face reconstruction accuracy, we propose a multi-attribute dynamic learning network (MADLNet) to recalibrate the label confidence for different facial attributes according to their sensitivities to the environment, which allows the model to adaptively refine their understanding of facial attributes without adding extra parameters. Specifically, we first randomly generate a number of hard samples with occlusion rates ranging from 0% to 60%. This creates challenging samples reflecting the different occlusion responses of different facial attributes. Then, we define sensitivities to occlusion for different facial attributes to recalibrate the confidence of facial multi-attribute labels so as to disentangle the optimization process. Finally, we calculate the multi-scale similarity between input samples and the created challenging samples to fine-tune label confidence, which addresses the misalignment problem between occlusion rates and semantic information caused by random occlusion operation. Extensive experiments conducted on the AFLW2000-3D and AFLW datasets effectively confirm the efficacy of our approach.