adpd-anonymous-review/adpd-review

Agile Drone Pose Dataset (ADPD)

ADPD is an indoor stereo benchmark for close-range agile drone trajectories. It provides synchronized stereo images, MoCap-derived 6-DoF annotations, official sequence-level splits, detector annotations, and camera calibration. It does not require LiDAR, point clouds, or dense scene reconstruction.

Dataset Summary

Split	Sequences	Stereo pairs	Images
Train	29	13,517	27,034
Validation	8	3,621	7,242
Test	4	1,700	3,400
Total	41	18,838	37,676

The split is defined over complete trajectory sequences. This avoids temporal leakage between adjacent frames from the same flight. Camera frames were captured at approximately 10 Hz. Each frame is temporally associated with the MoCap trajectory, with position and orientation evaluated from the bracketing source records using linear interpolation and quaternion SLERP, respectively. Released samples are at most 30 ms from the nearest source record and have a maximum bracketing interval of 100 ms.

Layout

data/{train,val,test}/sequence_NNN/
  left/*.jpg
  right/*.jpg
  annotations.csv
metadata/{train,val,test}.parquet
yolo/images/{train,val,test}/*.jpg
yolo/labels/{train,val,test}/*.txt
calibration/stereo/
calibration/world_to_camera/

The sequence CSV files are convenient for direct image-directory training. The split-level Parquet files provide a compact searchable index with paths relative to the repository root.

Annotation Fields

Field	Meaning
frame_id	Zero-based frame identifier within a trajectory
left_image, right_image	Relative paths to the stereo images
camera_timestamp	Camera system timestamp
sync_delta_ms	Distance to the nearest source MoCap record, in milliseconds
x_m, y_m, z_m	Drone position in the MoCap/world frame, in meters
qx, qy, qz, qw	Motive rigid-body orientation quaternion in XYZW order

The official world axes are defined as follows: x points toward the stereo camera, y points toward the right side of the camera image, and z points vertically upward. ADPD provides 6-DoF ground truth. The OTE benchmark reported with this release estimates the 3D position trajectory only.

Loading

from pathlib import Path
import sys

root = Path(".").resolve()
sys.path.insert(0, str(root / "scripts"))
from load_adpd import iter_samples

sample = next(iter_samples(root, "train"))
left_image, right_image = sample.open_images()
print(sample.position_m, sample.quaternion_xyzw)

The Parquet index can be loaded directly:

import pandas as pd

train = pd.read_parquet("metadata/train.parquet")

YOLO Detector Subset

The yolo/ directory contains 1,504 images with one drone class in standard YOLO format. Its sequence-level split contains 1,025 training, 238 validation, and 241 test images. This is the detector-training subset used by the Motion Tracker; it does not imply that every ADPD frame has a detection box.

Calibration

Stereo calibration uses a checkerboard with 16 by 11 inner corners and a 17 mm square size. Machine-readable files define intrinsics, distortion, left-to-right extrinsics, rectification matrices, units, and residuals. The OpenCV camera convention is x right, y down, and z forward.

The world-to-camera package provides a validated rigid transform derived from reviewed fixed-marker stereo correspondences. It defines the position transform as

p_left_camera = R_world_to_left_camera @ p_world + t_world_to_left_camera

and includes its inverse, the supporting correspondences, and stereo reprojection residuals. This supporting calibration does not change the official world-frame labels or evaluation protocol.

Validation

Run the complete local validation before use or upload:

python scripts/validate_release.py . --full-image-check
sha256sum -c checksums.sha256

Intended Use

ADPD supports stereo 3D position estimation, trajectory estimation, pose-aware analysis, detector training, and controlled geometric evaluation on unseen indoor flight trajectories from the official sequence-level split. Results on this split should not be interpreted as evidence of cross-camera, cross-site, or cross-platform generalization without additional evaluation.

License

ADPD is released under the Creative Commons Attribution 4.0 International License. Third-party software used to process the data retains its own license.

Citation

Citation metadata will be updated with the archival publication record.