3D reconstruction algorithms and systems at production scale.

About Me

I am a 3D reconstruction algorithm engineer focused on computer vision and photogrammetry. My work spans Structure from Motion (SfM), Multi-View Stereo (MVS), SLAM, point-cloud processing, mesh reconstruction, texture mapping, and multi-view geometry.

I work across algorithm design, validation, optimization, and C++/CUDA implementation, with a focus on accurate, efficient, and reliable reconstruction for large real-world scenes.

我是一名专注于计算机视觉三维重建与摄影测量的算法工程师，研究和实践涉及 SfM、MVS、SLAM、点云处理、网格重建、纹理映射与多视几何。

我的工作贯穿算法设计、验证、优化及 C++/CUDA 工程实现，关注大规模真实场景中三维重建的精度、效率与稳定性。

Current focus. 3D Gaussian Splatting and emerging end-to-end 3D reconstruction methods, including foundation models such as VGGT.

Research & Engineering Interests

Photogrammetry & 3D Reconstruction

Structure from Motion, multi-view stereo, aerial and oblique imagery, feature matching, bundle adjustment, and panoramic reconstruction.

SLAM & Multi-Sensor Fusion

Visual and LiDAR SLAM, LiDAR-camera calibration, sensor synchronization, lens correction, and spatial pose estimation.

Large-Scale Reconstruction Algorithms

Large indoor and aerial scenes, partition-and-merge reconstruction, CUDA optimization, memory efficiency, and reconstruction quality evaluation.

Point Clouds, Meshes & Rendering

Point-cloud filtering and registration, mesh reconstruction, texture mapping, level of detail, and reconstruction visualization.

Selected Projects

Representative work for reconstruction engineering roles

Large-scale reconstruction pipeline screenshot

Large-Scale Reconstruction & LOD 3D Tiles

Large scenes · Streamable reconstruction pipeline

A production-oriented reconstruction pipeline that organizes large indoor and mixed scenes into fixed-scale spatial tiles for bounded memory, parallel processing, mesh merging, texturing, simplification, and browser delivery.

Technical work
Absolute-scale tile pyramid; streaming point-cloud partitioning; seam-aware mesh merge; visibility and texture stages; two-phase LOD generation for 3D Tiles.
Engineering contribution
Unified reconstruction, floor handling, texture mapping, simplification, and delivery around one spatial key system so large scenes avoid global memory growth.

Tech: C++, point clouds, mesh reconstruction, texture mapping, LOD, 3D Tiles

Read case study

Indoor structured reconstruction result screenshot

Indoor Structured Reconstruction

Indoor geometry · Structural model generation

A geometry-first pipeline that converts noisy, multi-layer indoor meshes into clean structural models suitable for floorplans, measurements, and VR display.

Technical work
Mesh cleanup; plane extraction; wall and floor reasoning; room topology construction; graph optimization for closed structural geometry.
Engineering contribution
Moved reconstruction outputs from visual mesh data toward editable, measurable, application-ready indoor structure.

Tech: C++, mesh processing, plane extraction, topology, graph optimization

Read case study

Cyclops model refinement before and after screenshot

Cyclops Model Refinement

Panoramic depth · Production geometry cleanup

A refinement workflow for monocular panoramic depth reconstruction outputs, focused on making model predictions more consistent, cleaner, and usable in downstream geometry pipelines.

Technical work
Pose consistency checks; outlier filtering; point-cloud fusion; mesh cleanup; reconstruction validation from panoramic depth outputs.
Engineering contribution
Bridged learned depth outputs and production reconstruction requirements by improving stability and geometric quality.

Tech: panoramic depth, pose BA, point-cloud fusion, mesh cleanup

Read case study

CAD-to-floorplan region graph screenshot

CAD-to-Floorplan Parsing

CAD geometry · Structured spatial data

A DXF parsing pipeline that infers walls, windows, doors, and room regions from drawing geometry and topology instead of relying on unreliable CAD layer names.

Technical work
DXF entity parsing; linework cleanup; constrained triangulation; region graph construction; rule-based inference for architectural elements.
Engineering contribution
Converted noisy drafting data into structured floorplan outputs that can feed reconstruction, measurement, and product workflows.

Tech: DXF, computational geometry, triangulation, graph reasoning, floorplan JSON

Read case study

Panoramic geometry preprocessing screenshot

Panoramic Geometry Preprocessing

Single panorama · Spherical image geometry

A combined set of panoramic geometry modules for estimating gravity alignment and Manhattan directions before downstream reconstruction and spatial understanding.

Technical work
Cube projection; vertical-line extraction; spherical geometry; RANSAC; vanishing-point voting; constant-time Manhattan candidate scoring.
Engineering contribution
Packaged low-level geometric priors into robust preprocessing steps for panorama-based reconstruction workflows.

Tech: equirectangular panoramas, spherical LUT, vanishing points, gravity rectification, RANSAC

Gravity rectificationVanishing points

Software

Tools built for practical use

InsightAT

Command-line one-stop SfM pipeline for aerial triangulation. Simple interface, robust reconstruction.

C++ · SfM · MVS Code / Documentation

fastmvs

High-speed dense matching software. Significantly faster than conventional MVS pipelines while maintaining reconstruction quality.

C++ · CUDA · MVS Project / Documentation

Free PTD

Free airborne LiDAR ground filtering software. Progressive TIN densification algorithm with CPU parallel acceleration and block-based processing for large datasets.

C++ · LiDAR · TIN Project / Documentation

Technical Case Studies

Short engineering write-ups without full internal detail

  1. Indoor Structured Reconstruction

    Converts noisy multi-layer indoor meshes into closed, single-layer structural models for floorplans, measurements, and VR display.

    Mesh cleanup · Plane extraction · Room topology · Graph optimization

    CASE
  2. Large-Scale Reconstruction & LOD 3D Tiles

    Organizes city- and building-scale reconstruction as tiled, streamable stages with controlled memory, seams, floors, textures, and browser LOD.

    Tiling pyramid · Streaming mesh merge · Visibility · 3D Tiles

    CASE
  3. Cyclops Model Refinement

    Stabilizes monocular panoramic depth reconstruction by improving pose consistency, filtering outliers, and refining mesh generation.

    Panoramic depth · Pose BA · Point-cloud fusion · Mesh cleanup

    CASE
  4. CAD-to-Floorplan Parsing

    Turns noisy DXF linework into structured floorplan JSON without relying on layer names or manually curated drafting conventions.

    DXF parsing · Constrained triangulation · Region graph · Rule inference

    CASE
  5. Single-Panorama Gravity Rectification

    Estimates camera tilt from one panorama so downstream reconstruction can assume gravity-aligned imagery.

    Cube projection · Vertical lines · Spherical geometry · RANSAC

    CASE
  6. Panorama Vanishing Point Estimation

    Finds three Manhattan directions in equirectangular panoramas using spherical line evidence and constant-time candidate scoring.

    Vanishing points · Spherical LUT · Manhattan model · Robust fitting

    CASE

Experience & Education

3D Reconstruction Algorithm Architect
RealSee (贝壳-如视) · Beijing

3D Reconstruction Algorithm Engineer
Feima Robotics (飞马机器人) · Beijing

Co-founder · 技术总监
Infinite3D (北京无限界/景致三维) · Beijing

3D Reconstruction Team Lead
Xi'an Meihang Technology Development Research Institute (西安煤航技术发展研究院) · Xi'an

WebGIS Engineer
Heilongjiang Bureau of Surveying, Mapping & Geoinformation (黑龙江省测绘与地理信息局) · Harbin

M.S. Geographic Information Systems
Wuhan University (武汉大学) · Hubei

B.S. Resource Environment & Urban-Rural Planning
Northeast Petroleum University (东北石油大学) · Heilongjiang

Contact

For technical discussions, open-source collaboration, or reconstruction engineering work, please contact me by email.

huyang0909@gmail.com · github.com/huluoboge