SPIN-UV

SPIN-UV is a multimodal dataset for unstructured scene understanding in dense urban villages. It was collected from a motor-driven, human-steered single-track vehicle and pairs front-facing visual observations with frame-anchored riding-state signals. The dataset is intended to support semantic segmentation, RGB-D perception, state-conditioned traversability, temporal consistency, and motion-aware scene understanding in narrow, weakly structured urban-village corridors.

SPIN stands for Scene, Proprioception, Interaction, and Navigation. UV refers to the urban-village setting that anchors the dataset.

Dataset Components

The release contains two complementary parts.

`dataset_seg_split/`

This is the ready-to-train semantic-segmentation subset. It contains anonymized RGB images and COCO polygon annotations, split at sequence level into train, validation, and test partitions.

Each split folder contains PNG images and one COCO annotation file:

dataset_seg_split/
  train/
    *.png
    _annotations.coco.json
  val/
    *.png
    _annotations.coco.json
  test/
    *.png
    _annotations.coco.json

The split is sequence-level rather than image-level, so frames sampled from the same traversal do not cross train, validation, and test partitions.

`dataset_sequences/`

This is the processed raw multimodal sequence release. It contains 89 traversal sequences and 132,293 uploaded files after excluding per-sequence config_snapshot.json files.

Verified raw-sequence statistics:

A typical sequence is organized as:

dataset_sequences/
  sequence_YYYYMMDD_HHMMSS/
    events/
      events.csv
    frames/
      rgb/
      depth/
      depth_native/
      depth_preview/
      ir/
      frame_index.csv
      synced_frame_state.csv
    meta/
      camera.json
      session.json
    state/
      state.csv

The raw sequence folders include Intel RealSense D435i RGB-D and infrared streams, timestamp records, camera metadata, traversal metadata, event logs, and state signals such as speed, wheel RPM, steering angle, roll, pitch, yaw, gyroscope, accelerometer, and temperature fields.

Semantic Labels

The native semantic taxonomy contains 19 labels:

bicycle, bin, building, car, clutter, dog, person, pole, rider,
road, roadblock, scooter, sidewalk, sky, trafficsign, tricycle,
truck, vegetation, wall

The COCO export also contains a root category named CUVscapes as the Roboflow project category. The paper benchmark trains with the full native taxonomy and reports the main diagnostic mIoU on 13 analysis labels:

building, scooter, road, car, wall, clutter, sky, sidewalk,
vegetation, pole, person, tricycle, roadblock

The remaining six labels, bicycle, bin, dog, rider, trafficsign, and truck, remain in the annotations as supplemental context labels.

Quick Start

Download only the trainable semantic-segmentation subset:

hf download ruikle123/SPIN-UV \
  --repo-type dataset \
  --include "dataset_seg_split/**" \
  --local-dir SPIN-UV

Download the full raw sequence release:

hf download ruikle123/SPIN-UV \
  --repo-type dataset \
  --include "dataset_sequences/**" \
  --local-dir SPIN-UV

Read a COCO annotation file:

import json
from pathlib import Path

ann_path = Path("SPIN-UV/dataset_seg_split/train/_annotations.coco.json")
with ann_path.open("r", encoding="utf-8") as f:
    coco = json.load(f)

print(len(coco["images"]))
print(len(coco["annotations"]))
print([c["name"] for c in coco["categories"]])

Inspect the aligned RGB-D and state streams for one sequence:

import pandas as pd
from pathlib import Path

seq = Path("SPIN-UV/dataset_sequences/sequence_19700101_080403")
frame_index = pd.read_csv(seq / "frames" / "frame_index.csv")
synced_state = pd.read_csv(seq / "frames" / "synced_frame_state.csv")

print(frame_index.columns.tolist())
print(synced_state[["frame_id", "speed_mps", "steering_angle_actual_deg", "roll_deg", "pitch_deg", "yaw_deg"]].head())

Collection Platform

SPIN-UV was collected with a single-track vehicle platform operated by a seated human rider. The rear hub motor provides propulsion, while the human rider controls direction through the handlebar. The sensing stack records:

• RGB, aligned depth, native depth, depth preview, and infrared observations from an Intel RealSense D435i.
• Passive steering feedback.
• Rear wheel and hub-motor feedback.
• External IMU attitude and inertial measurements.
• Software timestamp alignment between visual frames and riding-state streams.

The current capture configuration stores RGB and depth at 1280x720. Temporal analyses should use the recorded timestamps in the CSV files rather than assuming a fixed frame rate.

Intended Uses

This dataset is designed for:

• RGB semantic segmentation in dense urban-village scenes.
• RGB-D semantic segmentation and multimodal feature fusion.
• State-conditioned traversability and obstacle/free-space reasoning.
• Temporal and motion-aware perception for single-track vehicles.
• Domain-transfer studies from structured-road datasets to narrow urban-village corridors.

Limitations and Responsible Use

The dataset focuses on a specific urban-village, low-speed, single-track traversal regime. It is not intended to represent all cities, all road conditions, all weather, or all vehicle platforms. Models trained on SPIN-UV should be evaluated carefully before use outside similar environments.

The semantic subset uses anonymized RGB images. The raw sequence release contains rich multimodal traversal data and should be used only for research and development purposes consistent with the dataset license. Do not attempt to identify people, riders, locations, or private attributes from the data.

Citation

If you use SPIN-UV, please cite the dataset and this repository. A paper citation will be added when the manuscript is public.

@dataset{spin_uv_2026,
  title = {SPIN-UV: A Multimodal Dataset for Unstructured Scene Understanding in Urban Villages},
  author = {SPIN-UV Authors},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ruikle123/SPIN-UV}
}

License

This dataset is released under the Apache 2.0 license.

📊 Structured Schema (Zero-Fabrication)

Estimated Rows: 868