LBG: A Local Reconstruction Method Based on Global Prior

Depth implicit function has been widely used as an effective method to represent 3D shapes. However, one drawback is that it is difficult to balance the accuracy and smoothness of the model. To resolve this problem, we propose a local reconstruction method based on global prior. Our idea is that most areas of the shape do not necessitate overly precise processing, thereby prompting us to segment only the intricate details. Our method fully utilizes global prior information to pinpoint the location of localized shapes, ensuring accurate and precise results. Due to the fact that our method does not need to consider the surrounding area for local shape, it can have arbitrary topological structures. To ensure seamless integration between local and global shapes, we used latent code to represent shapes as signed distances and blend global and local code in latent space. This approach effectively mitigated discrepancies and improved overall shape accuracy. Furthermore, our method can be learned without ground truth signed distances and point normals. Our approach has been shown to outperform existing methods when applied to both shape and scene datasets based on classical metrics.