CaRaCTO-3D Dataset

Camera + radar + motion-capture ground-truth data for extrinsic calibration between a camera and a 24 GHz FMCW radar, collected with a trihedral corner-reflector calibration target. This is the dataset behind:

• CaRaCTO: Robust Camera-Radar Extrinsic Calibration with Triple Constraint Optimization, published at ICPRAM 2024 (Best Industrial Paper Award).
• CaRaCTO-3D: From Camera-Radar Calibration to Scene Reconstruction, published in SN Computer Science 2025.

Dataset Summary

A single capture session: the camera+radar rig stayed fixed for the whole session while a trihedral corner-reflector target was placed at 40 different static positions in a room. At each position, camera, radar, and OptiTrack motion-capture data were recorded simultaneously, and the target was manually annotated in the camera image.

Dataset Structure

├── setup.png                            # reference photo of the room/rig setup
├── calibration/
│   ├── camera_intrinsics.json           # {"camera_matrix": 3x3, "dist_coeff": [5]}
│   ├── axis_convention.json             # the published coordinate frame convention
│   └── radar_rig_markers.json           # 4 rig-mounted mocap markers, published frame (fixed
│                                         # for the whole session — the rig never moved)
├── positions.json                       # per-position index: annotation/quality/alignment flags
└── Position_01/ ... Position_40/        # one directory per static target position
    ├── camera/
    │   ├── frame_00.jpg .. frame_NN.jpg  # full captured burst (count varies, 20-26 frames)
    │   └── frame_timestamps_ms.json
    ├── annotation.json                   # manual 6-point target annotation + PnP distance
    ├── radar.npz                         # radar detections for this position
    └── ground_truth.json                 # transform-corrected OptiTrack target marker positions

positions.json

One entry per position:

{
  "annotated": true,
  "valid_for_reconstruction": false,
  "num_camera_frames": 26,
  "num_radar_detections": 25,
  "frame_alignment": {
    "aligned": true,
    "camera_frame_range": [0, 25],
    "radar_detection_range": [0, 25],
    "note": null
  },
  "notes": null
}

• annotated: whether the corner-reflector target was successfully hand-annotated in the camera image for this position. 4 of 40 positions are not: Position_12, Position_16, Position_17, Position_35.
• valid_for_reconstruction: a hand-curated subset (28 of 40 positions) that achieved good camera/radar/depth correspondences for the 3D scene-reconstruction pipeline in the original research code. This is a research judgment call from the original authors, carried over verbatim, not something re-derived from the data.
• frame_alignment: see "Camera/radar frame alignment" below.

Position_XX/annotation.json

The corner reflector is a trihedral target annotated by 6 image points — an "outer" triangle and an "inner" triangle, one point pair per edge — whose corresponding lines all meet at the target's projected center:

{
  "annotated": true,
  "outer_triangle": [[x, y], [x, y], [x, y]],
  "inner_triangle": [[x, y], [x, y], [x, y]],
  "target_center": [x, y],
  "corner_edges": [[[x, y], [x, y]], [[x, y], [x, y]], [[x, y], [x, y]]],
  "distance_m": 4.074520093282576
}

corner_edges[i] is [outer_triangle[i], intersection_point], where intersection_point is where the line through outer_triangle[i]/inner_triangle[i] crosses the edge formed by the other two inner-triangle points — used to refine the target's 3D pose via solvePnP, giving distance_m. For the 4 unannotated positions, all fields except annotated are null.

Position_XX/radar.npz

range_m, velocity_mps, magnitude, angle_rad, noise    float32 (N,)   # per detection
amplitude_re, amplitude_im                             float32 (N,4) # per-antenna complex amplitude

N (the detection count) varies per position, typically 20-25. These are what the original calibration code actually used (averaging range_m and angle_rad across all detections in a position to get a single radar range/azimuth measurement).

The original raw capture (radar_data.pickle) also contained two extra per-position fields whose generation could not be attributed to any script found on the original processing machine (including the full git history of two other local forks of the processing code) — they were not used by any code in this repository, so they were dropped from this release rather than published as unexplained data.

Camera/radar frame alignment

Each position's camera burst (camera/frame_*.jpg) and radar detection list (radar.npz) come from independently-clocked acquisition loops that start/stop within a few cycles of each other, so their counts are usually equal or off by 1-2. There are no reliable per-detection timestamps on the radar side, so alignment assumes a constant frame rate per modality (not wall-clock timestamps) and trims the longer sequence's extra frames symmetrically from the start/end. positions.json's frame_alignment field records this per position:

• aligned: true — counts differ by at most 4; camera_frame_range/radar_detection_range give the trimmed, index-aligned window into each modality.
• aligned: false — counts differ by more than that (e.g. Position_10: 21 camera frames vs. only 4 radar detections; Position_30: 21 vs. 39). This is a real radar detection-count anomaly (track drop-outs or extra spurious detections), not a start/stop timing offset, so no automatic trim is applied — note explains why.

caracto.dataset.caracto_dataset.CaractoDataset.load_aligned(key) applies this automatically and raises for aligned: false positions.

Position_XX/ground_truth.json

{
  "target_markers_xyz": [[x, y, z], [x, y, z], [x, y, z], [x, y, z]],
  "target_center_xyz": [x, y, z],
  "description": "..."
}

The 4 individual corner-reflector target markers (mean over ~630 raw OptiTrack frames for this position) and the corresponding rigid-body centroid, both already centered on the radar and expressed in the published coordinate frame (see below). Only this corrected result is included, not the raw per-frame export.

OptiTrack Ground Truth & Coordinate Frames

Published frame: origin at the radar sensor position, X forward (away from the radar, into the scene), Y left, Z up — matching the convention used throughout the calibration/reconstruction code (azimuth = atan2(y, x), elevation along z), documented in calibration/axis_convention.json.

All ground truth in this release is already centered on the radar and expressed in this frame. radar_rig_markers.json gives the rig's own mocap marker positions as an independent sanity check — they cluster within ~5cm of the origin, as expected.

Known Limitations / Data Quality

• 4 of 40 positions have no manual camera annotation: Position_12, Position_16, Position_17, Position_35.
• 28 of 40 positions are flagged valid_for_reconstruction — a hand-curated "good correspondence" subset from the original research code. One quirk preserved verbatim: Position_35 is in this list despite having no annotation; this has no practical effect since annotation-based filtering is applied first by the calibration code.
• Two individual OptiTrack target markers were entirely untracked (mocap occlusion) for their whole position: Position_16 marker 1, Position_33 marker 2 — recorded as NaN in target_markers_xyz for that marker only; target_center_xyz (the rigid-body centroid) is unaffected since OptiTrack computes it from the other tracked markers.
• Raw radar ADC channel data (per-antenna chirp samples) was captured but is not included in this release — only the processed per-detection values are. It was unused by any of the original calibration/reconstruction code. Contact the authors if you need it for a specific research purpose.
• A duplicate copy of the original radar_data.pickle existed at two paths in the raw capture directory; investigation during this dataset's preparation found the two were not identical — one copy had one of its (since-dropped, unattributable) legacy fields overwritten with the OptiTrack ground truth value (for all 40 positions) and was missing the axis-convention field entirely, consistent with a later debug/analysis artifact rather than a genuine duplicate. This release is derived from the original (unmodified) copy, matching what the calibration code has always read.

Loading the Dataset

from caracto.dataset.caracto_dataset import CaractoDataset

# Local copy:
ds = CaractoDataset("/path/to/CaRaCTO-3D")
# Or directly from the Hub:
# ds = CaractoDataset(repo_id="dfki-av/CaRaCTO-3D")

sample = ds.load("Position_01")  # everything as published (full burst, full detections)
aligned = ds.load_aligned(
    "Position_01"
)  # camera frames + radar detections, index-aligned
single = ds.load_single(
    "Position_01"
)  # one representative measurement (frame 0 + averaged radar range/azimuth)

Without caracto installed, every file is a plain JSON/JPEG/NPZ file readable with standard tools (json.load, PIL.Image.open/cv2.imread, numpy.load).

Citation

@article{chamseddine2025caracto,
    title     = {CaRaCTO-3D: From Camera-Radar Calibration to Scene Reconstruction},
    author    = {Chamseddine, Mahdi and Rambach, Jason and Stricker, Didier},
    journal   = {SN Computer Science},
    volume    = {6},
    number    = {7},
    pages     = {822},
    year      = {2025},
    publisher = {Springer},
}

@inproceedings{chamseddine2024caracto,
    title        = {CaRaCTO: Robust Camera-Radar Extrinsic Calibration with Triple Constraint Optimization},
    author       = {Chamseddine, Mahdi and Rambach, Jason R and Stricker, Didier },
    year         = 2024,
    booktitle    = {Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods - ICPRAM},
    pages        = {534--545},
    organization = {INSTICC},
}

Acknowledgement

This research was partially funded by the European Union as part of the project HumanTech (Grant Agreement 101058236) and the Federal Ministry of Education and Research (BMBF) of the Federal Republic of Germany as part of the research project COPPER (Grant Number 01IW24009).

License

This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.