Multi-Level Pixel-Aligned Implicit Function for High- Resolution 3D Human Digitization

Abstract

Current strides of image dependent 3 dimension human outline estimation have progressed due to remarkable strides in depiction capabilities facilitated by deep NN. Despite the strides made in real-world applications, existing methods still fall short in generating reconstructions that match the intricate details often found in the original images. We posit that this deficiency primarily arises from the clash between two competing demands: accurate predictions necessitate extensive contextual information, while precise predictions hinge on higher resolutions. Owing to the limitations in current hardware memory, prior techniques have leaned towards utilizing low-resolution images to encompass broader spatial context, resulting in less precise or lower-resolution 3D estimations. To overcome this hurdle, we have devised multi layered algorithm that undergoes endwise training. At rough level, model comprehensively processes the entire image at a reduced resolution, emphasizing holistic reasoning. This coarse level furnishes essential context to a finer level, which focuses on estimating highly intricate geometries by scrutinizing higher-resolution images. Our research demonstrates a substantial enhancement in performance compared to prior methodologies, showcasing the superior capabilities of our approach.