A 3D Gaussian Splatting reconstruction of my day-to-day desk at Dartmouth. The desk was captured on a phone, 38 s of 60 FPS FHD video. The standard pipeline used for processing is described below.
46.2 MB · loads on demand
camera: ready
Since the capture is quite dense, I downsampled the number of temporal frames — I only used every fourth frame in my capture. That's 38 s × 60 FPS / 4 ≈ 570 frames. To capture the camera extrinsics and intrinsics, I ran the standard COLMAP pipeline.
For Gaussian Splatting, I used the 3DGS-MCMC pipeline with a max cap of 6,000,000 splats with initial 30,000 iterations. After training, I used a different train/test split, to finetune further for 5,000 iterations twice.
The final scene was rendered in your browser using mkkellogg/gaussian-splats-3d.
Given the initial scene was over 1 GB, I compressed it using the provided KSPLAT format to ~50 MB.