Improving Implicit Geometry Learning with Multiscale Tri-Plane Feature Encoding

In recent years, with the rapid progress of deep learning in computer graphics, substantial efforts have been devoted to the exploration of diverse data structures for the representation of neural implicit fields. To boost the efficiency in learning and rendering the neural implicit fields, a novel tri-plane representation has been proposed to encode the neural radiance fields or geometry fields. Although appealing results are achieved by using the tri-plane feature encoding for the task of novel view synthesis, the potential for using tri-plane feature encoding to represent the intricate geometry is not well studied. To improve the current performance on using tri-plane feature encoding for implicit geometry learning, we propose a novel multiscale tri-plane feature encoding scheme to characterize features in different scales. Specifically, features from multiscale tri-planes are concatenated together to form a multiscale feature for learning the implicit geometry fields such as SDF. Quantitative and qualitative experiments are conducted to show the effectiveness of our multiscale tri-plane feature encoding scheme.