Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
1.92k
1.92k
End of preview. Expand in Data Studio

Kojori Canyon FPV Photogrammetry

One flight of a foam fixed-wing FPV plane over the canyon south-east of Kojori, Georgia, on 2025-11-22: 241 video keyframes, the flight controller's GPS track, and the reference outputs of a full reconstruction (camera poses, a dense cloud in metres, a 2 m height grid).

It is the worked example for the fpv-photogrammetry toolkit and the data behind the write-up Measuring a Canyon from FPV Video. Source footage: https://youtu.be/s8RxP2BK_d0.

Why it is a useful test case:

  • The flight was not planned for mapping: a forward-looking action camera with electronic stabilisation and a rolling shutter, no survey grid, no ground control points.
  • Five passes from different directions and heights, so matching has to bridge rotation and scale.
  • An independent GPS track gives scale and a check that reprojection error cannot give.
  • One pass (c43) reconstructs as a mirror image with a normal reprojection error.

Contents

path what size
images/ 241 keyframes, 1920×1080 JPEG, named <clip>_t<seconds in the clip>.jpg 99 MB
mini/images/ 59 of them (clips c40 and c42, two passes over the same slope) for a quick SfM run 26 MB
clips.csv per clip: source file, offset from arm, segment used, frame counts 1 KB
telemetry/gps_5hz.csv GPS fixes, 5 Hz 1.4 MB
telemetry/pose_25hz.csv EKF position and attitude, 25 Hz 3.8 MB
telemetry/terrain_1hz.csv SRTM-derived terrain height under the aircraft, 1 Hz 0.1 MB
telemetry/meta.json arm events, home position, clock fit, GPS summary 4 KB
reference/sparse/0/ COLMAP sparse model (binary): 241 images, 136 979 points 60 MB
reference/geo/georef.json similarity transform from model units to metres, time lag, residuals 1 KB
reference/geo/cams_enu.csv camera centres in local east/north/up metres 22 KB
reference/geo/cloud_enu.npz dense cloud in east/north/up metres: xyz float32, rgb uint8, 8 004 328 points 120 MB
reference/geo/web/ 2 m height grid (dsm.png), orthophoto (ortho.jpg), meta.json 2.8 MB

Not included: the source video (4K, 120 fps, tens of GB) and the raw flight log.

Quick start

git clone https://github.com/format37/fpv-photogrammetry && cd fpv-photogrammetry
pip install -r requirements.txt huggingface_hub

examples/kojori.sh fetch     # download this dataset into ./kojori
examples/kojori.sh geo       # no GPU, under a minute: GPS fit + height grid from the reference outputs
examples/kojori.sh mini      # GPU: sparse SfM on the 59-frame subset, about 3 min
examples/kojori.sh full      # GPU: everything from the 241 frames, about 2.5 h on an RTX 4090

geo must reproduce reference/geo/georef.json and reference/geo/web/dsm.png exactly. mini and full run COLMAP, which is not bit-reproducible: expect numbers close to the ones below.

Without the toolkit:

import numpy as np, pandas as pd

cloud = np.load("reference/geo/cloud_enu.npz")        # xyz in metres (east, north, up), rgb 0..255
gps = pd.read_csv("telemetry/gps_5hz.csv")            # t_since_arm_s, lat, lon, alt_gps_msl, ...
cams = pd.read_csv("reference/geo/cams_enu.csv")      # name, t_arm, used, e, n, u

The cloud is unfiltered and has far outliers (haze, horizon). Crop it to the area of interest: the reference height grid covers east −900…1500 m, north −2000…100 m.

How it was captured

  • Airframe: ZOHD Altus, a 980 mm twin-motor V-tail foam wing.
  • Autopilot: Matek H743, ArduPlane 4.6. Automatic takeoff, then FBWA. Two u-blox M10 GPS receivers at 5 Hz, a barometer, EKF3 at 25 Hz.
  • Camera: DJI O4 Pro in the nose on a tilt mount, looking forward and slightly down almost all the time. Electronic stabilisation on, rolling shutter. Video 3840×2160, 119.88 fps, HEVC 10-bit, about 130 Mbit/s.
  • Flight: 27.4 min armed, 31 km, up to 663 m above the launch point, median ground speed 19.6 m/s.
  • Keyframes: one per about 2.5 % of image shift (sparse optical flow), the sharpest of each window, area-downscaled to 1920 px. clips.csv lists the segment of each clip that was used.

Camera intrinsics from the bundle adjustment (COLMAP OPENCV model, 1920×1080): fx 800.6, fy 801.9, cx 960, cy 540, k1 −0.1137, k2 0.0100, p1 0.0001, p2 −0.0008. The camera was never calibrated on a target; the focal length prior from the DJI metadata is 805.8 px.

Time and coordinates

Frame time. The camera starts recording on arm, and the recording is split into files (clips.csv: offset_s is the start of each file in seconds after arm). For a frame:

t_since_arm = offset_s(clip) + t in the file name + lag,   lag = 1.15 s

The lag is the camera's start-up delay. It was not measured: it is the value that minimises the fit of the SfM camera centres to the GPS track. It can only be found when the passes fly in different directions. For passes on one course, such as the two in mini/, a time lag looks like an along-track shift and the fit absorbs it.

Telemetry time base. t_since_arm_s is true seconds since the arm event. The flight controller clock ran 0.104 % fast (1.7 s over the flight); this is removed with a linear fit of log time to the GPS fix epochs (18 ms rms residual). utc is GPS time minus 18 leap seconds.

Coordinates. lat, lon are WGS84 degrees. alt_gps_msl is GPS altitude above mean sea level. The reference outputs use a local east/north/up frame with its origin at 41.6510441, 44.7295113, 1116.98 m. That is a pre-arm origin record from the log, about 1 m from the launch point and 6.3 m below it (launch at arm: 41.651046, 44.729524, 1123.3 m). So up + 1116.98 is height above sea level.

Telemetry columns.

  • gps_5hz.csv: one row per fix of GPS #2, the receiver the autopilot used. gps_speed in m/s, course in degrees from true north, vz_down in m/s positive down, hacc / vacc in metres, fix 3 = 3D, 4 = 3D with SBAS. roll, pitch, yaw, alt_rel, baro_alt are interpolated from the 25 Hz data to the fix time.
  • pose_25hz.csv: EKF3 position and attitude. alt_msl is the EKF altitude (barometer-driven; it drifts about 5.5 m against GPS altitude by landing), alt_rel the height above home.
  • roll, pitch, yaw: degrees, airframe attitude in the NED frame, Euler ZYX. Roll is positive right wing down, pitch positive nose up, yaw 0–360 from true north. The camera mount angle relative to the airframe is not known.
  • terrain_1hz.csv: terrain_msl is the height from ArduPilot's terrain database (SRTM-derived, 100 m grid); terrain_logged_msl is the logged value, shifted by ref_offset so that the terrain matches the aircraft altitude at arm; height_above_terrain as computed on board.

Reference results

stage result
Sparse SfM 241 of 241 frames in one model, 136 979 points, mean reprojection error 0.74 px, mean track length 10.7
GPS fit, 223 cameras (c43 left out) lag +1.15 s, scale 162.548 m per model unit, 3.03 m rms
GPS fit, each pass on its own 0.7–1.5 m rms
Dense MVS and fusion 8.0 million points
Height grid 2 m cells, 1200×1050; 78 % of the modelled area is measured, the rest is interpolated and flagged
mini/ subset, sparse SfM 59 of 59 frames, 42 567 points, mean reprojection error 0.67 px, focal length 803 px, GPS fit 3.4 m rms at the lag of the full set

Processing times on an RTX 4090: features and matching 2 min, mapper 31 min, PatchMatch stereo 89 min, fusion 3 min.

Known limitations

  • Mirrored pass. c43 (18 frames, a high pass over distant terrain) sits about a kilometre from its GPS positions in the reference sparse model, while its reprojection error is a normal 1.3 px. It is kept in images/ and in reference/sparse/0/ on purpose, and excluded from the reference georeferencing and cloud.
  • Drift between passes. The model is aligned to GPS after the reconstruction with one similarity transform. GPS is not used inside the bundle adjustment, hence 3 m globally against about 1 m per pass.
  • Surface, not ground. Forest is reconstructed at canopy height. Deep shadow and haze leave holes.
  • Video, not stills. The 4K frames carry little detail beyond 1920 px (0.13 bit per pixel after compression, electronic stabilisation resamples every frame).
  • Dense resolution. The reference dense stage ran on frames undistorted to 1920×911 at a focal length of about 637 px, an earlier default of the toolkit; the current default keeps 800 px.

License and attribution

CC BY 4.0. Please credit "format37, Kojori Canyon FPV Photogrammetry" and link to this page or to the write-up.

Downloads last month
-