Learning a Multimodal 3D Face Embedding for Robust RGBD Face Recognition

Machine vision will play a significant role in the next generation of IR 4.0 systems. Recognition and analysis faces are essential many vision-based applications. Deep Learning provides thrust for advancement visual recognition. An important tool recognition tasks is Convolution Neural networks (CNN). However, 2D methods machine suffer from Pose, Illumination, Expression (PIE) challenges occlusions. The 3D Race (3DFR) very promising dealing with PIE certain degree occlusions suitable unconstrained environments. data highly irregular, affecting performance deep networks. Most Face models implemented research aspect rarely find complete 3DFR application. This work attempts to implement end-to-end robust pipeline. For this purpose, we CuteFace3D. face model trained on most challenging dataset, where state-of-the-art had below 95% accuracy. accuracy 98.89% achieved intellifusion test dataset. Further, open world unseen domain adaptation, embeddings learning using KNN. Then FR pipeline RGBD RealSense D435 depth camera. With KNN classifier k-fold validation, 99.997% set registered users. proposed method early fusion four-channel input found be more has higher benchmark