建筑信息模型（BIM，Building Information Modeling）是建筑行业数字化转型中的重要一环，其能够实现建筑的全生命周期管理。 然而在实际的工程应用中，现场的工程进度很难直接反映在BIM模型中，因此BIM模型的更新往往只能依靠人工手动翻模，这导致BIM应用的成本高、效率低。 因此，通过智能化手段将现场采集的点云数据直接转换为 BIM 模型的智能翻模技术（Scan-to-BIM），对于实现现场与 BIM 的无缝对接具有极为重要的意义。 Scan-to-BIM中重建的对象主要分为两类，一类是建筑的主体结构，例如墙梁柱板等参数化构件，另一类是建筑的非参数化构件，例如桌椅、沙发等，这些非参数化构件的重建对于Scan-to-BIM的完整性至关重要。 本文针对Scan-to-BIM中非参数化对象的重建问题，设计并实现了一套场景理解与非参数化对象重建方法，本文的主要贡献包括以下三点： 在场景理解阶段，本文提出了一种面向实例重建的场景理解模型，该模型在基于软分组的实例分割算法的基础上，通过残差编码对语义实例的实例包围盒进行优化，在进行实例分割的同时计算产生实例包围盒，提升了准确性的同时为后续的重建提供了更多的信息。同时本文采用点云配准算法对实例重建结果和原始点云进行匹配，进一步提高场景理解的准确性。 在实例重建阶段，本文提出了一种面向场景内残缺实例点云的综合重建模型，该方法针对不同的实例类别分别采用低多边形重建模型和基于量化特征卷积的隐式重建模型，从而提升了不同类别实例的重建效果。此外本文还提出了一种基于场景内对象间自相似性的重建模块，该模块能够利用场景内对象间的自相似性，对实例重建的结果进行融合，从而提升了实例重建的效果。 本文设计并实现了 Scan-to-BIM 中非参数化对象的重建系统，该系统已成功与主体结构重建系统集成，实现了 Scan-to-BIM 的完整流程。此外，该系统还以插件的形式被集成于 BIM 建模软件 Revit 中，通过 Revit 软件可以直接调用本文提出的语义实例重建算法，实现基于场景的语义实例重建，使得 Scan-to-BIM 方法更加完整。 实验测试表明本文实现的场景实例重建模型相较于现有方法有着一定提升，在可视化效果上，本文方法在场景中存在多个相似对象的场景中，重建结果更加完整精细。

Building Information Modeling (BIM) is an essential component in the digital transformation of the construction industry, as it enables comprehensive lifecycle management of buildings. However, in practical engineering applications, updating BIM models solely through manual re-modeling due to the difficulty in directly reflecting on-site progress in the model, results in high costs and low efficiency of BIM applications. Therefore, intelligent means are needed to seamlessly integrate the field data with BIM, leading to the development of the Scan-to-BIM technique, which converts point cloud data collected on-site into BIM models. This technique has significant implications for the development of BIM. Reconstruction in Scan-to-BIM can be broadly categorized into two types: one is parametric components such as walls, beams, columns, and ceiling that constitute the main structure of the building; And the other is non-parametric instances such as tables, chairs, sofas, etc. The reconstruction of these non-parametric components is critical to the integrity of Scan-to-BIM. To address the reconstruction of non-parametric instances in the Scan-to-BIM technique, this paper proposes and implements a scene understanding and non-parametric instance reconstruction method. The main contributions of this work are as follows. In the scene understanding stage, a scene understanding model for instance reconstruction is proposed that optimizes the semantic instance bounding box by residual encoding based on a soft-grouping instance segmentation algorithm. The instance bounding box is calculated simultaneously with the instance segmentation, thereby improving accuracy and providing more information for subsequent reconstruction. Additionally, the instance reconstruction results are used to align the original point cloud, further improving the accuracy of scene understanding. In the instance reconstruction stage, a comprehensive reconstruction model for incomplete instance point clouds in the scene is proposed, which employs a low-polygon reconstruction model and a vector quantization-based convlolution DIF reconstruction model for different instance categories, thus improving the reconstruction performance of different categories of instances. Furthermore, a reconstruction module based on the scene-similarity between objects in the scene is proposed to fuse the instance reconstruction results, further improving the reconstruction performance. A non-parametric instance reconstruction system for Scan-to-BIM is designed and implemented in this work, which has been successfully integrated with the main structural reconstruction system, realizing the complete Scan-to-BIM process. This system has been integrated into the BIM modeling software Revit in the form of a plugin, and the proposed semantic instance reconstruction algorithm can be directly invoked through Revit software, enabling scene-based semantic instance reconstruction and enhancing the completeness of the Scan-to-BIM technique. Experiments show that the scene instance reconstruction model proposed in this work significantly improves upon existing methods. In terms of visualization, the proposed method produces more complete and refined results in scenes containing multiple similar objects.