PointRecode：基於 LLM 的複雜幾何體點雲至程式碼逆向工程系統

Abstract

從三維點雲逆向重建可編輯 CAD 模型，是 CAD 重建與參數化設計自動化中的重要問題。點雲可由網格表面、B-Rep 實體或生成式模型輸出經採樣取得，作為不同幾何來源的共同輸入介面；搭配大型語言模型，則可進一步生成具備參數入口與操作結構的 CadQuery 程式碼。然而，既有點雲至 CAD 方法多集中於草圖—拉伸（sketch-extrude）資料分布，對包含倒角、圓角、殼體、旋轉、掃掠、放樣與布林運算等廣泛操作（broad-operation）的程式碼重建，仍缺乏系統性研究。本研究提出 PointRecode，一個針對 Zero-to-CAD-1M（ZTC）廣泛操作分布訓練與評估的點雲至 CadQuery 程式碼生成框架，並探討三項核心組件：用於擷取局部幾何線索的 LocalPooling 編碼器、依程式複雜度排序的課程訓練，以及針對高風險操作與 B-Rep 執行失敗的雙階段 AST 修復管線。

實驗結果顯示，局部幾何編碼在廣泛操作重建中至關重要。在 ZTC validation single-shot 評估中，LP-Max 編碼器相較於 Fourier-only 基線，將 Voxel IoU 由 32.64% 提升至 45.51%（增加 12.87pp），顯示編碼器架構對幾何重建精度的影響顯著大於點雲解析度。進一步結合複雜度排序課程訓練後，最佳單次推理模型 Exp G 在 ZTC 上達到 57.90% Voxel IoU 與 84.6% 執行率。

為提升執行穩定性，雙階段 AST 執行修復管線在第一階段進行高風險操作 fallback，第二階段於 B-Rep 執行失敗時進行選擇性修飾移除。在 Exp G 的 1000 筆樣本評估中，此管線將執行率由 70.9% 提升至 90.2%，Voxel IoU 由 46.4% 提升至 58.0%。跨實驗失敗模式分析顯示，revolve 為最常見的高風險操作，而 loft、sweep 及 polygon 則因樣本稀少構成瓶頸；AST 修復管線結合失敗診斷，能有效針對高風險與稀有操作進行復原，證明執行修復對整體幾何重建品質具有實質增益。

在推論時增強設定下，使用多溫度的實驗（10-Temp TTO）使 Exp G 於 ZTC 子集上達到 75.32% Voxel IoU 與 100% 執行率。此結果大幅超越以八張多視角影像為輸入的微調 Qwen3-VL-2B 基準模型（61.33% IoU），而對照組 CAD-Recode 搭配 TTO 則達 88.60% IoU，顯示 PointRecode 在點雲輸入下具備極佳的推論穩定性與競爭力，但幾何精度仍有提升空間。本研究為從點雲重建可編輯、可執行且具操作結構的 CAD 模型奠定了實證與診斷基礎。

Reverse engineering editable CAD models from 3D point clouds is crucial for design automation. However, existing methods focus on sketch–extrude workflows, leaving the problem of broad-operation CAD reconstruction (e.g., chamfer, fillet, revolve, sweep, loft) understudied. We propose PointRecode, a point-cloud-to-CadQuery generation framework evaluated on Zero-to-CAD-1M (ZTC). We investigate three core components: a LocalPooling encoder for capturing local geometric features, complexity-ordered curriculum learning, and a two-stage AST execution repair pipeline.

Our experiments highlight the importance of local geometric encoding. In single-shot ZTC evaluation, the LP-Max encoder increased Voxel IoU from 32.64% to 45.51% compared to a Fourier-only baseline (+12.87 pp), showing that encoder architecture impacts geometric accuracy more directly than point resolution. Incorporating complexity-ordered curriculum learning further optimized the model (Exp G) to achieve 57.90% Voxel IoU and an 84.6% execution rate.

To improve execution stability, the two-stage AST repair pipeline applies fallback logic to high-risk modifiers and selectively strips failing operations upon B-Rep failures. For Exp G, this pipeline increased the execution rate from 70.9% to 90.2% and Voxel IoU from 46.4% to 58.0%. Failure analysis shows that revolve is the most consistent high-risk operation, while loft, sweep, and polygon pose bottlenecks. The execution-aware repair successfully mitigates these failures, improving geometric quality beyond simple syntax post-processing.

Under test-time optimization via multi-temperature experiments (10-Temp TTO), Exp G achieved 75.32% Voxel IoU and a 100% execution rate on a ZTC subset. This outperforms the fine-tuned Qwen3-VL-2B multi-view image baseline (61.33% IoU). Meanwhile, CAD-Recode with TTO achieved 88.60% IoU, showing PointRecode's competitiveness and potential for further refinement. This work establishes a solid foundation for reconstructing editable, executable CAD models from point clouds.