Edge Prediction for Roof Wireframe Reconstruction with Transformers
Researchers have developed a novel Transformer-based method for reconstructing 3D roof wireframes from sparse point clouds. This approach, inspired by DETR, dynamically subsamples input data and fuses it with semantic and Gestalt features. The system achieved a Hybrid Structure Score of 0.6476 on the "HoHo 22k" dataset, securing second place in the S23DR Challenge 2026. AI
IMPACT Introduces a novel Transformer architecture for 3D reconstruction, potentially improving scene understanding in computer vision.