@wu_peng follow-up to A Set of DJI Drone Photos Fails to Generate an Orthophoto (closed now).
We ran your 99 M3D images (odm_datasets.zip) with your exact command (without --gcp) and with variations:
| options | reconstructed points | orthophoto |
|---|---|---|
your command (--sfm-algorithm triangulation, feature-quality low, fast-orthophoto) |
1 329 | broken (only a 56 × 52 m patch) |
same + --auto-boundary |
1 213 | empty |
same but without --sfm-algorithm triangulation |
19 442 | complete, 258 × 800 m, 10 cm, 2.8 min |
Both orthophotos at the same scale
Cause: triangulation does not estimate the camera orientations, it trusts the gimbal angles in the EXIF. In 23 of your images (at the start, 0004–0033, and the end, 0095–0100) the camera was really tilted 11–51° away from the recorded −90°. The points triangulated from those images land far outside the site; in your run the model bounds became ±16 km (Model bounds x: -15933 → 10126), and the orthophoto for that area needs 558 GB – hence the memory error.
Fix: use the default sfm-algorithm (incremental) and keep --auto-boundary. The rest of your command can stay as it is.