Research map / Point clouds / Point cloud classification

ArcPro: Architectural Programs for Structured 3D Abstraction of Sparse Points

CVPR 2025 · Point clouds › Point cloud classification · approach: Point-language and open-vocabulary 3D · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ArcPro, a DSL of architectural programs plus a 3D-conv encoder and transformer decoder that maps sparse building point clouds to programs, which are compiled into low-face-count mesh abstractions.

Problem

Recovering clean, structured 3D abstractions of buildings from highly sparse, noisy, non-uniform, or incomplete point clouds (e.g., SfM-derived), where primitive-detection pipelines fail.

Approach

Represent buildings as programs in a hierarchical architectural-tree DSL, synthesize unlimited program-point cloud pairs via procedural generation, and train an autoregressive encoder-decoder to predict programs from sparse points, with an FSM-based token masking to guarantee syntactic validity and a learning-free compiler to produce meshes.

Key results

• On the SfM test set, ArcPro reaches HD 0.0154 and LFD 3873 with 36 faces and 100% success rate, versus ProxyRecon HD 0.0243 and LFD 4364.
• On the sparse sampled test set, ArcPro reaches HD 0.0219 and LFD 4932, outperforming the best baseline ProxyRecon (HD 0.0256, LFD 5340).
• In a 110-participant user study, ArcPro was preferred in 76.7% of 20 cases over five baselines; it processes 1,090 buildings (2.92 km2) in about 37 seconds on one 4090 GPU.

Measured effects reported in the paper

Surprising findings

• Top-3 sampling reduced output quality rather than improving diversity, so the model is run with top-1 sampling.

Limitations

• Lacks precision for detailed structures because training relies only on next-token prediction loss.
• Sparse inputs can admit multiple valid solutions, and the method outputs only a single top-1 solution.
• The DSL currently covers only prismatic/stacked layers and cannot represent curved surfaces or sloped roofs.

Experimental setup

Datasets: Bing Maps building footprints UrbanBIS UAV SfM point clouds (270 buildings)

Benchmarks: SfM point cloud test set (270 buildings) UrbanBIS-derived sparse point cloud test set (1,038 buildings)

Metrics: Hausdorff Distance Light Field Distance success rate vertex/face/plane counts user preference

Baselines: PolyFit KSR ProxyRecon CA BSP-Net

Builds on: BSP-Net CA (cuboid abstraction) procedural grammar modeling Marching Cubes

Closest papers by meaning

• BuildAnyPoint: 3D Building Structured Abstraction from Diverse Point Clouds (CVPR 2026 · similarity 0.89)
• SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.85)
• 3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code (NeurIPS 2026 · similarity 0.85)
• LinNet: Linear Network for Efficient Point Cloud Representation Learning (NeurIPS 2024 · similarity 0.85)
• Draw Step by Step: Reconstructing CAD Construction Sequences from Point Clouds via Multimodal Diffusion. (CVPR 2024 · similarity 0.84)
• Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.84)