Drone photos
Overlapping aerial photos from a DJI drone circling a two-storey shingle-roof building. Heavy overlap (60–80%) is what lets the math work.
source: open datasetSparse reconstruction
SIFT features matched frame-to-frame, cameras solved by incremental bundle adjustment (COLMAP / pycolmap). This is the actual output you can spin below.
computed on this VMDense mesh + texture
Multi-view stereo densifies the cloud into a textured mesh — but it needs a GPU and hours of compute, so the textured model below is the dataset author's own reconstruction.
borrowed referenceThe raw material — 8 of the 61 photos
Shot from a drone on an orbital flight path around the building. Each frame overlaps its neighbours — that's the fuel for the reconstruction.
What we computed — the sparse point cloud
Every dot is a real 3D point triangulated from matched features across the photos. Drag to orbit, scroll to zoom. It ran here on 2 CPUs with no GPU.
The finished look — textured reference model
The dataset author's dense reconstruction of the same building — textured mesh from multi-view stereo. Ours stopped at the sparse stage: dense MVS needs a beefy GPU machine.
HONEST LABELS — WHAT RAN WHERE
On this VM (just now): SIFT feature extraction, sequential matching and incremental bundle adjustment over 61 downsampled photos → 61 registered, 33,516 3D points. No GPU. That point cloud above is 100% ours.
Borrowed from the dataset: the original 61 full-resolution photos and the textured .glb mesh (the author's own dense multi-view-stereo reconstruction). Our VM can't run dense MVS.