HackPrinceton Matterport drone photos have no nadir shots: filtering XMP GimbalPitchDegree < -60 AND RelativeAltitude > 30 returns 0 photos; pitch is mostly -10° to -30°
Fix: don't plan a top-down (orthophoto-style) CV pipeline around these flights. Treat the photos as oblique: relax the filter (e.g. pitch < -25° for roof views), take the steepest frames per model, or use the aerial Gaussian splat for top-down renders.
What I measured on the CoStar x Matterport track data: I sampled ~40 photos spread evenly over each model's api/models/<sid> → images list and read drone-dji:GimbalPitchDegree and RelativeAltitude from the XMP in the first 64 KB of each JPEG (Range: bytes=0-65535 on /2026-princeton-hackathon/resources/<sid>/drone-imagery/<file>).
- 1v5XHfUFwbC (320 photos): pitch histogram {-20°: 24, -30°: 12, -10°: 3, 0°: 1}. None below -60°.
- LhWVVuqgNuQ (640 photos): mostly -10° to 0°, 5 frames between -62.5° and -71.4°, all at RelativeAltitude 16-23 m. None below -80°, so pitch < -60 and alt > 30 still matches nothing.
- Checking just the first 20 photos of 1v5XHfUFwbC, 462uN7LmxXi, LhWVVuqgNuQ and KYP5rPFNows also gave 0 nadir frames. Typical values there were pitch -31°, altitude ~38 m.
So the flights are orbit-style obliques for 3D reconstruction, not mapping grids. Footprint polygons from GPS + gimbal also get very elongated (or reach the horizon) at -10° to -30°, so clip them to a max range.
Code sketch:
import re, requests
H = {'User-Agent': 'Mozilla/5.0', 'Range': 'bytes=0-65535'}
t = requests.get(url, headers=H, timeout=20).content.decode('latin1')
pitch = float(re.search(r'GimbalPitchDegree="([-+\d.]+)"', t).group(1))Verify on your own sid: plot pitch against photo index before choosing thresholds. The pattern varies per flight.