Periocular recognition has emerged as a particularly valuable biometric identification method in challenging scenarios, such as partially occluded faces due to COVID-19 protective masks masks, in which face recognition might not be applicable. This work presents a periocular recognition framework based on deep learning, which automatically localises and analyses the most important areas in the periocular region. The main idea is to derive several parallel local branches from a neural network architecture, which in a semi-supervised manner learn the most discriminative areas in the feature map and solve the identification problem solely upon the corresponding cues. Here, each local branch learns a transformation matrix that allows for basic geometrical transformations (cropping and scaling), which is used to select a region of interest in the feature map, further analysed by a set of shared convolutional layers. Finally, the information extracted by the local branches and the main global branch are fused together for recognition. The experiments carried out on the challenging UBIRIS-v2 benchmark show that by integrating the proposed framework with various ResNet architectures, we consistently obtain an improvement in mAP of more than 4% over the “vanilla” architecture. In addition, extensive ablation studies were performed to better understand the behavior of the network and how the spatial transformation and the local branches influence the overall performance of the model. The proposed method can be easily adapted to other computer vision problems, which is also regarded as one of its strengths.
In this paper, we propose multi-phase fine-tuning for tuning deep networks from typical object recognition to sign language recognition (SLR). It extends the successful idea of transfer learning by fine-tuning the network’s weights over several phases. Starting from the top of the network, layers are trained in phases by successively unfreezing layers for training. We apply this novel training approach to SLR, since in this application, training data is scarce and differs considerably from the datasets which are usually used for pre-training. Our experiments show that multi-phase fine-tuning can reach significantly better accuracy in fewer training epochs compared to previous fine-tuning techniques
In this demo, we build upon VRS-NeRF to provide an implementation for real-time variable rate shading (VRS) in virtual reality (VR). Here, we demo a recent advancement of Neural Radiance Fields (NeRF) merging multiple pixels into one reducing the overall rendering time. This system allows the real-time visualization of small to medium NeRF scenes. Further, we demonstrate the applicability of VRS-NeRF for 2D applications and further show its generalizability by providing two additional methods for the determination of the shading rate and an adaptive method for adjusting the step size during ray marching.