Jamesbass/bigsur-360-colmap

Big Sur Cliff — 360° Drone COLMAP Reconstruction + Gaussian Splat

A coastal bluff in Big Sur, California, captured in a single 10-minute flight of a DJI Avata 360 drone (8K equirectangular video), reconstructed with COLMAP's native spherical camera support and trained into a 3D Gaussian Splat with LichtFeld Studio.

View it in your browser

Open ColmapView → Load URL → paste:

https://huggingface.co/datasets/Jamesbass/bigsur-360-colmap/resolve/main

You get the sparse point cloud, the drone's camera trajectory, the source panoramas, and the trained splat overlaid in one coordinate frame.

Contents

Path	Description
sparse/0/	COLMAP 4.1.0 sparse reconstruction (binary, rigs/frames format). 773 registered EQUIRECTANGULAR cameras, 360k points, 0.85 px mean reprojection error
images/	773 source panoramas, 7680×3840 equirectangular JPEG (sharpest frame per 0.75 s of video)
splats/bigsur_v2.ply	Trained gaussian splat, 10M gaussians (SH degree 1 export for browser viewers)
bigsur_v2_full_quality.sog	Same splat, full SH degree 3, compressed SOG format

Pipeline

1. DJI Studio export: 8K equirectangular MP4 (color-restored from D-Log M)
2. Sharpest-frame extraction: 1 frame per 0.75 s (Laplacian variance per chunk)
3. SfM: COLMAP 4.1.0 panorama_sfm.py --pano_render_type spherical — native equirectangular camera model, sequential matching, incremental mapper (819/819 frames registered; takeoff frames removed with image_deleter)
4. Point filtering: reprojection error > 2 px, track length < 3, far outliers removed
5. Training: LichtFeld Studio PR #1369 (--undistort expands each panorama into 12 × 90° virtual pinhole views → 9,276 training views), MRNF strategy, SH 3, mip filter, bilateral grid, 10M gaussian cap, 45k iterations
6. Final quality: PSNR 31.64 / SSIM 0.923 on held-out views

Moving ocean water reconstructs as a soft averaged surface (expected for any static-scene method).