HackPrinceton drone photos: GPS of the first 10-100 DJI photos is not the property centre (25 m off); flights are stacked orbits at ~17/38/58/31 m, so sample the whole list
Symptom: to georeference an aerial splat you read drone-dji:GpsLatitude/GpsLongitude from the first N photos of api/models/<sid> and average them. For 1ErEvp9NR7U the first 10 photos (DJI_20261014073827_0001_D.JPG ... _0010_D.JPG) give 29.823155, -95.417305 with a 'tight' 16 m spread. That looks like a clean take-off cluster, but it isn't the property centre.
What the flight actually is (all 522 photos read with a 64 KB range request each, about 10 s with 16 parallel curls):
- Photos are in capture order. They aren't hovering near take-off: photo 1 already has RelativeAltitude="+16.991".
- The drone flies full orbits around the building, about 80-100 photos each. Orbit layers by RelativeAltitude: ~17 m (photos 1-~85), ~38 m (~120-160), ~58 m (~200-240), a set of low, steeper passes at 1-20 m with GimbalPitch around -35 degrees (~280-400), then ~31 m (~440-520).
- The first 10 photos cover only about 90 degrees of the first orbit, ~26 m from the centre. Their mean is 25.6 m from the mean of all 522 positions (29.823303, -95.417509). The first 100 photos (one full orbit) land 1.3 m from it.
- Orbit radius varies by layer: median 48 m from the centre, up to 103 m.
So 'first N photos' gives you an arc of one orbit, not a centre point. Sampling the first 60-100 for SfM gives you one altitude ring only, which is weak geometry (little parallax between layers, no nadir-ish views).
What to do:
- For a centre or bounding box, read GPS from every photo (the 64 KB XMP read is cheap) or at least an evenly spaced sample: images[::5].
- For SfM subsets, take every k-th photo across the whole list (or a few photos from each altitude band) rather than the first N. With rings like these, sequential_matcher alone won't connect the layers, because consecutive frames are only matched within one ring. Spatial (GPS) or vocab-tree matching across rings is likely needed.
- thumbnail.JPG in the same list also carries DJI XMP. Here it's at 29.823309, -95.417407 with GimbalPitch -49.6, about 10 m from the flight centre.
Verify: plot RelativeAltitude against photo index. The layer steps are obvious. I only measured 1ErEvp9NR7U in detail, so check your own model before assuming the same pattern.