MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202621058351 A) filed by Indian Institute Of Technology Gandhinagar on May 07, 2026, for System And Method For Self-Supervised 3d Point Cloud Completion.
Inventors include Kumari, Seema; Kumar, Preyum; Mandal, Srimanta; and Raman, Shanmuganathan.
The application for the patent was published on July 03, 2026, under issue no. 27/2026.
Abstract: ABSTRACT SYSTEM AND METHOD FOR SELF-SUPERVISED 3D POINT CLOUD COMPLETION The present disclosure provides a system(100) and method(200) for self-supervised completion of three-dimensional(3D) point clouds. The system(100) comprises an input module(101) that receives an incomplete point cloud, a patch generation module(102) that generates local geometric patches, and a feature encoder module(103) extracts discriminative feature representations. The system further includes a multi-head self-attention module(104) captures long-range contextual dependencies, and a latent representation module(105) configured to generate a compact latent representation. A decoder module(107) reconstructs a dense and complete point cloud using MLP-based decoding and coarse-to-fine residual refinement. During training, a contrastive learning module(106) performs symmetric contrastive latent alignment based on InfoNCE loss between partial and reconstructed point clouds, while a loss optimization module(108) applies region-aware Chamfer Distance and normal variance loss for robust geometric reconstruction and surface smoothness. The system operates without requiring paired complete ground-truth datasets and enables efficient and reliable completion of irregular and incomplete point cloud data.
Disclaimer: Curated by HT Syndication.
