3D building wireframe reconstruction via edge point regression

Despite significant progress in point cloud–based 3D building reconstruction, the effectiveness of existing methods remains constrained by four major bottlenecks. These include the scarcity of high-quality annotated datasets, the pervasive incompleteness and noise in real-world point clouds, the substantial memory consumption incurred by the edge search space, and the lack of adequate geometric constraints in network designs, which often causes loss functions to fail in capturing the underlying 3D structural regularities. To address these challenges, we pre-train a self-supervised learning (SSL) framework on large-scale unlabeled point clouds to learn generic geometric representations, and subsequently fine-tune it using only a limited set of annotated roof samples for 3D wireframe reconstruction. The key contribution of this work is the proposal of a new SSL-based framework that utilizes geometry aware edge point identification to facilitate wireframe reconstruction and avoid missing corners, thereby mitigating issues related to dataset quality and computational complexity. Specifically, the framework first leverages a pretrained SSL-based feature extractor to generate robust point-wise representations under sparse and imperfect inputs. Building on these features, an edge point–guided regression module predicts the parametric attributes of roof wireframes. To effectively supervise the training of this regression module, we formulate a joint loss function that optimizes two complementary objectives: (1) minimizing the Euclidean distance between input points and their nearest line segments, and (2) reducing the distance between candidate edge points and their projections on the reconstructed model. Coupled with bipartite edge matching and an edge-aware non-maximum suppression strategy, the proposed method ensures geometric validity while reducing computational redundancy. Extensive experiments across three datasets of varying geometric complexity and scale (Dense, Building3D, and the large-scale Tallinn City dataset) demonstrate the framework’s effectiveness. The SSL-based pretrained model attains 97% corner precision and 91% edge precision for wireframe reconstruction on the Building3D dataset, surpassing conventional baselines and approaching the performance of fully supervised methods. Remarkably, even under limited supervision, specifically utilizing only 50% partially labeled data during fine-tuning, the model maintains strong accuracy with 86% corner precision and 67% edge precision, underscoring its advantages over current SSL-based approaches.