All projects

Native CUDA · Native WebGPU · SfM

Find the cameras. Build the scene.

Working system

Splatfiction Mapper

Native CUDA and native WebGPU camera alignment and sparse geometry—the reconstruction stage before Gaussian training.

Aerial dataset source: TerraSky3D / A 0002. Reconstructed from overlapping drone frames in Splatfiction Mapper.

Scope & contribution

What I built.

Mapper estimates camera relationships and sparse scene structure from overlapping images through native CUDA and native WebGPU implementations. The work joins feature matching, geometric verification and bundle adjustment into a reconstruction stage designed to feed later Gaussian training.

My role
Independent developer
Technology
Native CUDA · Native WebGPU · SfM · Bundle adjustment
01

Camera alignment

Built the pipeline that estimates relative camera motion from image correspondences.

02

Geometry checks

Added verification stages that reject weak or inconsistent matches.

03

GPU implementations

Developed native CUDA and native WebGPU paths for refining the reconstructed cameras and scene.

Workflow overview

Feature matches
Camera alignment
Sparse reconstruction

Continue exploring

Part of a wider visual-computing practice.

See how the reconstruction, rendering and simulation work connects across projects.

All projects

Start a conversation

Let’s talk
about the work.

Engineering, visual computing, or something between the two.

Get in touch