Built the indoor large-scene reconstruction system behind station-scanned spaces such as Beike VR house tours, museums, and large venue digitization.

• Designed the full tiled parallel algorithm framework so reconstruction, mesh processing, texturing, simplification, and LOD generation can run beyond single-machine memory limits.
• Implemented topology-consistent block mesh merging, robust hole filling, irregular multi-floor segmentation, automatic ceiling hiding, semi-global color balancing, textured simplification, LOD strategy, and cloud Docker deployment.

Tech: C++, point clouds, mesh reconstruction, topology repair, texture optimization, LOD, Docker

Read case studyRealsee examples

Designed a geometry pipeline that recovers room structure from a given indoor mesh and reconstructs a regularized structural mesh.

• Built the full framework from scratch: space partitioning into a cell complex, visibility-based inside/outside reasoning, and graph-cut optimization.
• Independently implemented the complex cell-complex geometry algorithms, room partitioning, door recognition, and structure reconstruction modules.

Tech: C++, mesh processing, cell complex, visibility, graph cut, structural reconstruction

Read case study

Refined monocular panoramic depth outputs into more stable point-cloud and mesh geometry for downstream reconstruction pipelines.

• Added pose consistency checks, outlier filtering, point-cloud fusion, and mesh cleanup.
• Bridged learned depth predictions with production geometry quality requirements.

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

Read case study

Designed an automatic DXF/DWG parser to convert non-standard manually drawn CAD floorplans into a unified structured floorplan representation.

• Handled inconsistent layers, drawing styles, line types, units, blocks, and geometry conventions across legacy CAD files.
• Independently designed and implemented algorithms for layer parsing, inside/outside classification, room partitioning, door/window recognition, door block recognition, furniture filtering, and floorplan data generation.

Tech: DXF/DWG, computational geometry, triangulation, region graph, door/window recognition, floorplan JSON

Read case study

Developed panorama preprocessing modules for gravity alignment and Manhattan direction estimation before reconstruction and spatial understanding.

• Implemented cube projection, spherical line reasoning, RANSAC, vanishing-point voting, and fast Manhattan scoring.
• Packaged geometric priors as reusable preprocessing steps for panorama-based workflows.

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

Gravity rectificationVanishing points