Berzelius255/motive-loop-closure

Motive Loop Closure Dataset

A fleet-scale collection of dashcam clips and GPS telemetry capturing genuine driving loops — routes where a vehicle returned to within metres of a prior location after driving away. Designed as a benchmark and training resource for loop closure detection, visual place recognition (VPR), and visual odometry (VO) drift characterisation.

Dataset Description

Summary

Each entry in this dataset is a loop sequence: a pair of 3-minute dashcam clips from the same vehicle — one captured on the vehicle's first pass through a location (origin), and one captured on the vehicle's return (revisit), anywhere from minutes to several hours later. Ground-truth revisit distance is measured by GPS.

The dataset covers urban and mixed driving across a commercial freight fleet in Accra, Ghana, collected via the Motive fleet management platform.

Source data

• Vehicles: 5 vehicles (anonymised as v101, v102, v103, v106, v113)
• Collection period: September 2026
• Camera: Front-facing dashcam, 1920×1080, 29.97 fps
• GPS: ~1 Hz fixes from the Motive vehicle location API

Loop detection method

Candidate loop pairs were identified by scanning GPS traces for fix pairs where:

• Temporal gap > 30 s (vehicle had moved away)
• Haversine revisit distance < 50 m
• Path length ≥ 300 m (filters depot/stationary false positives)
• Speed > 1 m/s at both endpoints

Candidates were ranked by revisit_m / duration_s (lower = tighter return over longer drive). The top 33 candidates with confirmed video availability were recalled and are included here.

Dataset Structure

README.md
extract_frames.py          # Run locally to extract full frames from raw videos
metadata/
  loop_pairs.parquet       # One row per loop sequence — key statistics
  metadata.jsonl           # One record per clip — GPS endpoints + loop pair reference
data/
  v102_20260909_loop01/
    origin/
      preview/             # 5 evenly-spaced JPEGs for quick visual inspection
        000001.jpg ... 000005.jpg
      telemetry.csv        # timestamp_utc, lat, lon, speed_mph, heading_deg
    revisit/
      preview/
      telemetry.csv
  v113_20260909_loop01/
    ...
videos/
  v102_origin_20260909_183645_front_facing.mp4   # raw 3-min clips (180 s, 29.97 fps)
  v102_revisit_20260909_231842_front_facing.mp4
  ...

Extracting full frames

The dataset ships 5 preview frames per clip to keep download size small. To extract all frames at any frame rate:

# Clone or download the dataset, then:
python3 extract_frames.py              # 2 fps (default), all sequences
python3 extract_frames.py --fps 5     # 5 fps
python3 extract_frames.py --sequence v102_20260909_loop01 --fps 10

Requires ffmpeg on PATH. Frames land in data/<sequence>/<role>/frames/<fps>fps/.

metadata/loop_pairs.parquet schema

Column	Type	Description
sequence_id	string	e.g. v102_20260909_loop01
vehicle	string	Anonymised vehicle ID
origin_clip_start_utc	ISO datetime	Start of the origin clip
revisit_clip_start_utc	ISO datetime	Start of the revisit clip
revisit_gps_m	float	GPS distance between loop endpoints (metres)
loop_duration_s	float	Elapsed seconds between origin and revisit
origin_video	string	Filename in videos/
revisit_video	string	Filename in videos/

metadata/metadata.jsonl schema

One record per clip (2 per loop sequence: origin + revisit).

{
  "clip_id":         "v102_20260909_loop01_origin",
  "sequence_id":     "v102_20260909_loop01",
  "role":            "origin",
  "vehicle":         "102",
  "clip_start_utc":  "2026-09-09T18:36:45+00:00",
  "clip_duration_s": 180,
  "video_path":      "videos/v102_origin_20260909_183645_front_facing.mp4",
  "telemetry_path":  "data/v102_20260909_loop01/origin/telemetry.csv",
  "preview_frames":  ["data/v102_20260909_loop01/origin/preview/000001.jpg", "..."],
  "gps_start":       {"lat": 5.6037, "lon": -0.1870},
  "gps_end":         {"lat": 5.6041, "lon": -0.1873},
  "loop_partner": {
    "clip_id":         "v102_20260909_loop01_revisit",
    "revisit_gps_m":   0.53,
    "loop_duration_s": 17007
  }
}

Note: Visual similarity scores between origin/revisit clip pairs are not included in v1.0. GPS distance between loop endpoints is the ground-truth signal. Community contributions of VPR similarity annotations (NetVLAD, DINOv2, SuperPoint+SuperGlue, etc.) are welcome.

Intended Uses

Use case	What to use
Loop closure detection benchmark	loop_pairs.parquet — pair origin/revisit sequences, test if your detector fires
Visual place recognition evaluation	Frame pairs across origin / revisit roles; GPS as ground truth
VO drift characterisation	Full clip sequences with GPS; measure accumulated drift at closure
Self-supervised descriptor training	GPS-supervised positive pairs (same location, hours apart)
HD map staleness detection	Same route, different dates — detect scene changes

Limitations

• Single camera: Front-facing only. No stereo, no lidar.
• GPS accuracy: ~1–10 m typical for commercial fleet GPS. Revisit distances below ~2 m should be treated as "sub-metre" rather than exact.
• No visual similarity ground truth: Frame-level correspondence labels are not provided in v1.0.
• Weather / time-of-day: Not labelled; inferred from timestamps.
• v110 footage: All recalls for vehicle 110 timed out on the Motive platform and are absent from this dataset.

Citation

If you use this dataset, please cite:

@dataset{motive_loop_closure_2026,
  title   = {Motive Loop Closure Dataset},
  year    = {2026},
  note    = {HuggingFace Datasets},
  url     = {https://huggingface.co/datasets/<username>/motive-loop-closure},
  license = {CC BY 4.0}
}

License

Creative Commons Attribution 4.0 International (CC BY 4.0)