Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
mesh
mesh

RAVIDATA: RAdar, VIsual and inertial DATAset

60 GHz FMCW radar, phone video and IMU, recorded together. Licence: CC BY-NC 4.0 (attribution, no commercial use). Tools to record, align and analyse it: https://github.com/zRafaF/ravidata (make data downloads this dataset; make pod makes the maps on a GPU).

Recordings of a Possumic RS6130 60 GHz FMCW radar (MRS6130-P1812 board, 1 TX / 3 RX) next to a phone camera (Samsung SM-S936W, 1440 Γ— 1440, 30 fps, Log profile) and the phone's IMU (Sensor Logger, 100 Hz), for research on obstacle detection with a radar on a moving, hand-held platform (bachelor thesis, UTFPR, 2026). Recorded on 2026-11-13 in Brazil, indoors and outdoors.

  • Radar still: an empty scene, then placed objects, then a person walking (tags 0 chest, 1 back) and a broom (tags 20, 21, 50, 51) swept through the field of view, two backgrounds; a pilot with three chairs.
  • Radar + phone carried along an indoor and an outdoor course with AprilTags on the floor (10, 12,
    1. as metric checkpoints.
  • Every scene is recorded with three radar configurations, one after the other, filmed by one continuous video.

Takes

take radar parts recordings video description
static_pass1 still S0, S1, S2 9 0.05 GB Radar still, indoor lab: empty scene (S0) β†’ objects placed (S1) β†’ person walking (tags 0 chest, 1 back) + broom with four tags (20, 21, 50, 51) swept through the field of view (S2). Background 1.
static_pass2 still S0, S1, S2 9 0.04 GB As static_pass1, another background / objects.
indoor_course carried D1 3 0.08 GB Radar + phone carried along an indoor course; floor tags 10, 12, 17; one pass of the course per radar config.
outdoor_course carried D2 3 0.40 GB As indoor_course, outdoors.
pilot_chairs still chairs, walk 5 0.02 GB Pilot: radar still, three chairs with tags 11/13/15 at ~2 m, then a person walking (two 1D, two 2D, one 1D recording). Muted video (may contain conversation).

Radar configurations

config mode range bin max range frame period (fps) cube per frame
tcc-1d-40mm 1D range (4 chirps averaged on chip) 4 cm 5.12 m 30 ms (33.3) 1 Γ— 3 RX Γ— 128
tcc-2d-40mm-12fps 2D range-Doppler (16 bins, Β±2 m/s, 0.25 m/s) 4 cm 5.12 m 80 ms (12.5) 16 Γ— 3 RX Γ— 128
tcc-2d-40mm 2D range-Doppler (16 bins, Β±2 m/s, 0.25 m/s) 4 cm 5.12 m 50 ms (20.0) 16 Γ— 3 RX Γ— 128

The chip runs the range FFT (and in 2D the Doppler FFT); each frame is the complex cube after those FFTs. The 2D configs lose 5–15 % of the frames on the SPI link (mostly single frames); every frame carries its own host time (t_host), so use it rather than assuming a constant rate.

Files

README.md               this card
rig.json                mounting: radar rotation, antenna pairs, camera offset
apriltags.json          tag36h11 IDs, printed sizes (mm) and where each used tag was ("used")
calibration/
  camera.calib.json     intrinsics from the checkerboard clip ("display" block = the upright 1440 Γ— 1440 video)
  camera.calib.png      corners used + an undistorted frame (faces blurred)
  checkerboard.mp4      the calibration clip (faces blurred)
takes/<take>/
  video.mp4             phone video: upright, faces blurred, original frame times and sound; HEVC 10-bit, Log
  imu/                  Sensor Logger export: one CSV per sensor (gyroscope, accelerometer, gravity, orientation, ...;
                        device id removed from Metadata.csv)
  take.json             what was recorded, every radar recording, the clock alignment ("sync")
  radar/NN_<part>_<config>.h5
      cube          (frames, Doppler bins or intervals, 3 RX, range bins) complex64
      t_host        (frames,) host clock, Unix seconds
      frame_index   (frames,) the radar's own frame counter (gaps = lost frames)
      range_m       (range bins,)        v_axis (Doppler bins,) m/s, 2D only
      attrs: config_json (the radar configuration), derived, recording_json (part, beeps, loss, lead-in)
derived/
  maps-v1/              feed-forward 3D maps of the two carried takes (GPU pod, `make pod` in the code repository)
    <model>/<take>/win_NNN.npz, run.json   per window of 64 (or 48) keyframes: poses, depth, confidence, K
    frames_<take>.json  the keyframes given to the models (video frame, time, flow, sharpness), K, tone curve
    fused/<model>/<take>/  global map made on a CPU (`make fuse`): poses.npz (pose graph + ICP, per keyframe
                        camera-to-world), graph.json, fused.npz / fused.ply (TSDF points)
    logs/, metrics/     every log and running metric of the pod (GPU, time, memory, energy)
  static-depth-v1/<model>/<take>/   monocular metric depth of the still takes (CPU; Depth Pro `depthpro`,
                        Depth Anything V2 Small `da2s`), scaled on the AprilTags: .npz depth, .jpg frame, .ply, summary.json

The derived files are the exact inputs of the published analysis (results/analysis-v2 in the code repository). fused/ and static-depth-v1/ can be remade with make fuse / make static-depth, but not bit for bit (a re-run of the pose graph on another machine moved the keyframe poses by 5 cm median, up to 0.8 m), so use the uploaded ones to reproduce the published numbers.

Parts: S0 empty scene, S1 objects placed, S2 things moving (person, broom), D1 / D2 the indoor / outdoor course, chairs / walk in the pilot. The first recording of a take starts with a sync lead-in (lead_in_s, 8 s): a hand swipe or a tilt of the rig to the floor, not part of the scene.

AprilTags

tag36h11; size = edge of the black square. Only these IDs were placed (the others in apriltags.json were printed but not used):

IDs on size
0, 1 the person: 0 chest, 1 back 150 mm
10, 12, 17 the floor 90 mm
11, 13, 15 the chairs 90 mm
20, 21, 50, 51 the broom 55 mm (20, 21), 40 mm (50, 51)

Time alignment

All radar frames are on the host clock (t_host). The recorder beeps (1.2 kHz) once when a recording starts and twice when it stops; the beeps are found in the video's sound and matched to the recordings' beep times (take.json β†’ sync):

t_host(video frame) = video_a * t_video + video_b + recommended_video_offset_s
t_host(IMU sample)  = t_phone + imu_lag_s + recommended_video_offset_s

recommended_video_offset_s is the remaining delay between the logged beep time and the video. For the still takes it comes from radar-vs-video motion (recommended_because says which recordings); for outdoor_course it is the default 0.56 s (measured on the pilot take), which the radar floor return agrees with (+0.4 .. +0.6 s). For indoor_course the default was wrong: since v1.1 the offset is βˆ’2.44 s (the video was 3.0 s late), measured from the radar floor return against the camera height, the radar range against the map and tag ranges, and the radar Doppler against the walking speed from the map (details and the old value in take.json β†’ sync β†’ recommended_because, corrections). The beep fit itself is good to about one video frame; check the offset on your use.

Camera

The video is the phone's Log profile (BT.2020, flat): convert it (e.g. a Log-to-Rec.709 LUT or a tone curve) before using vision models. Intrinsics: calibration/camera.calib.json β†’ display (fx β‰ˆ fy β‰ˆ 1281 px, centre β‰ˆ (724, 712), k1 0.075, k2 βˆ’0.065), valid for the upright 1440 Γ— 1440 video. Autofocus was not locked; the focal length is known to about Β±1.5 %.

Known limitations

  • Radar height above the ground and the camera's dy / dz offsets were not measured (estimate them from the ground return and the IMU pitch, or from the video).
  • The pilot (pilot_chairs) and the calibration clip are muted (the pilot may contain conversation); their clock alignment was computed from the original sound.
  • One pass per config per course: configurations are compared on similar, not identical, motion.
  • No lidar or motion-capture ground truth: positions come from AprilTags (and from vision models you run).
  • The live radar angle shown during recording was uncalibrated; calibrate the RX phase offsets before using angles. The code repository fits them on the walking person of the still takes (azimuth: about 3Β° median error between takes); elevation stays approximate (the targets span only a few degrees).

Privacy

Faces are blurred (YuNet detector on every frame, boxes kept Β±8 frames, pixelated and blurred). The Sensor Logger device id is removed; no location is stored. The calibration clip and the pilot are muted. Elsewhere the published sound contains only the beeps: band-passed to 1.0–1.4 kHz (zero-phase) and kept Β±0.1 s around each beep found in the original, silence elsewhere. The clock alignment in take.json was computed from the original videos (full sound, identical frames and frame times); the beep onsets in the published sound match the originals within a few ms.

Versions

version date change
v1.0 2026-11-14 recordings, calibration, derived/maps-v1 (pod)
v1.1 2026-11-15 indoor_course video offset 0.56 β†’ βˆ’2.44 s (take.json, with the old value in sync.corrections); outdoor_course recommended_because reworded (value unchanged); derived/maps-v1/fused and derived/static-depth-v1 added; this card: derived files, time alignment, versions

hf download DoninhaD/ravidata --repo-type dataset --revision v1.0 gets the dataset as it was before v1.1 (the inputs of the code repository's results/analysis-v1, apart from fused/ and static-depth-v1, which were made locally then).

Loading

import json, h5py, av
take = json.load(open("takes/indoor_course/take.json"))
rec = take["recordings"][0]
with h5py.File("takes/indoor_course/" + rec["file"]) as f:
    cube, t_host, r = f["cube"][:], f["t_host"][:], f["range_m"][:]
s = take["sync"]
with av.open("takes/indoor_course/video.mp4") as c:
    for fr in c.decode(video=0):
        t = s["video_a"] * fr.time + s["video_b"] + s["recommended_video_offset_s"]   # host clock
        ...
Downloads last month
214