A Lightweight Geometric Consistency Enhancement for One-Stage Architectural Line Segment Detection
Zhiying Yang, Yichen Zong, Y HU, Yuanhan Ou · IEEE Access · 2026
Architectural line segment detection is an important geometric primitive extraction task for roof structure vectorization, indoor floorplan understanding, and downstream 3D reconstruction. Fully convolutional one-stage detectors such as F-Clip are attractive because they directly regress line centers, offsets, lengths, and orientations from dense heatmaps without expensive candidate-line verification. However, when applied to architectural scenes, the original one-stage line representation still faces two practical limitations. First, the independent supervision of length and orientation may lead to endpoint drift, where a small angular error on a long roof ridge, eave, wall boundary, or other structural line can be amplified into a large endpoint displacement. Second, dense heatmap responses often produce multiple near-duplicate line segments around the same physical edge, making structural post-processing essential for stable vectorization. To address these issues, this paper proposes a lightweight geometric consistency enhancement for one-stage architectural line segment detection. Built upon the efficient F-Clip representation, the proposed enhancement introduces endpoint-space supervision during training and structure-aware duplicate suppression during post-processing. Specifically, we propose endpoint vector supervision (EVS), a parameter-free auxiliary loss that reconstructs endpoint half-vectors from the predicted line length and orientation and applies a permutation-invariant endpoint-space constraint at positive line-center locations. EVS explicitly couples length and orientation regression in endpoint space without introducing additional learnable parameters or changing the inference architecture. During inference, we revisit structural NMS for architectural line detection and further analyze an angle-aware variant (AA-sNMS), which introduces an optional angle gate to examine the effect of directional consistency in endpoint-distance-based duplicate suppression. This analysis shows that the baseline structural NMS provides the main duplicate-suppression benefit, whereas the additional angle gate in AA-sNMS has only marginal and dataset-dependent effects. Experiments are conducted on RoofMapSet for remote-sensing roof structure extraction and on Structured3D for indoor structural line detection. On RoofMapSet, EVS improves the F-Clip baseline from 66.99/73.40/75.86 to 67.62/73.57/75.98 in sAP5/sAP10/sAP15, and combining EVS with structural NMS further improves the scores to 69.27/75.77/78.29, with the auxiliary endpoint-derived mAP ${}^{J}$ increasing from 35.04 to 37.00. On Structured3D, the proposed enhancement increases sAP5/sAP10/sAP15 from 54.27/63.08/66.43 to 56.49/65.67/69.34. Additional long-line endpoint error analysis shows that EVS brings larger relative improvements on long structural lines, supporting its intended role in reducing endpoint displacement. These results demonstrate that the proposed geometric consistency enhancement can improve one-stage architectural line segment detection without increasing the complexity of the underlying F-Clip representation.