Dataset Card for motor_two_wheel_rider

This is a FiftyOne dataset with 324 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/motor_two_wheel_rider")

# Launch the App
session = fo.launch_app(dataset)

MOTOR Dataset Card (FiftyOne Subset)

Dataset Details

Dataset Description

MOTOR (MOtorized TwO-wheeler Rider) is the first large-scale, multi-view, multimodal dataset dedicated to understanding two-wheeler rider behavior in dense, unstructured traffic conditions typical of the Global South. The full dataset comprises 1,629 annotated sequences (~25 hours) from 16 riders collected across diverse traffic scenarios in India.

This repository contains a subset of the MOTOR dataset imported into FiftyOne format for easy exploration and analysis. Each clip in the FiftyOne dataset contains 4 synchronized camera views (front-mounted, helmet-mounted, rear-mounted, and eye-tracker), organized as grouped samples for easy multi-view analysis. The dataset captures both conventional riding behaviors (going straight, turns, lane changes) and unconventional behaviors (weaving through traffic, obstruction avoidance, violations) with rich annotations including legality labels, traffic context, GPS trajectories, gyroscope data, vehicle speeds, and rider gaze patterns.

Key Features:

• 🎥 Multi-view video: 4 synchronized camera angles per clip

• 🏍️ 12 riding behaviors: Conventional and unconventional maneuvers

• ⚖️ Legality labels: Legal, illegal, or unspecified per traffic rules

• 📍 GPS routes: Frame-level coordinates and trajectory polylines

• 👁️ Gaze heatmaps: Attention patterns from eye-tracking glasses

• 🛣️ Rich context: Road type, lane count, traffic density, dividers, etc.

• Curated by: Varun Paturkar et al., IIIT Hyderabad

• Paper: MOTOR: A Multimodal Dataset for Two-Wheeler Rider Behavior Understanding (ICRA 2026)

• Project Page: https://varuniiith.github.io/MOTOR-Dataset/

• Hugging Face: https://huggingface.co/datasets/varunpaturkar/MOTOR

• License: Research use only (as specified by authors)

Dataset Sources

• Repository: https://github.com/varuniiith/MOTOR-Dataset
• Paper: https://arxiv.org/abs/2605.22550
• Project Page: https://varuniiith.github.io/MOTOR-Dataset/
• Hugging Face: https://huggingface.co/datasets/varunpaturkar/MOTOR

Uses

Direct Use

This subset is designed for:

• Quick exploration of the MOTOR dataset structure
• Prototyping analysis pipelines
• Understanding the data format before downloading the full dataset
• Testing FiftyOne workflows with grouped multimodal data

The full dataset is designed for:

• Rider behavior recognition: Train models to classify 12 riding maneuvers
• Legality prediction: Predict whether maneuvers comply with traffic rules
• Attention modeling: Analyze rider gaze patterns during different behaviors
• Multimodal fusion: Combine video, gaze, and telemetry for improved predictions
• Safety research: Study dangerous behaviors and near-collision events
• Traffic analysis: Understand two-wheeler behavior in dense traffic conditions
• ADAS development: Build advanced driver assistance systems for motorcycles

The FiftyOne format enables:

• Quick visual exploration with the FiftyOne App
• Spatial queries on GPS data
• Temporal queries on events and behaviors
• Frame-level analysis of gaze and telemetry
• Multi-view synchronized playback

Out-of-Scope Use

• Real-time inference: Clips are short (~1-20 seconds) and pre-segmented
• Autonomous driving: Dataset focuses on rider behavior, not scene understanding
• Non-Indian traffic: Traffic patterns are specific to dense, unstructured Indian roads
• Four-wheeler analysis: Dataset is specific to two-wheelers (motorcycles, scooters)

Dataset Structure

FiftyOne Organization

The dataset is organized as a grouped dataset with 4 slices per clip. You can switch between camera views in the FiftyOne App using the group slice selector, or programmatically by setting dataset.group_slice:

• front - Front-mounted camera (default) - includes GPS routes and telemetry
• helmet - Helmet-mounted camera
• rear - Rear-mounted camera
• eye_tracker - Eye tracker with rider's POV - includes gaze heatmaps

Sample-Level Fields

Each sample contains the following fields:

Field	Type	Description
clip_id	String	Unique clip identifier (e.g., "01_042")
video_id	Integer	Full ride video ID
camera	String	Camera view: front, helmet, rear, eye_tracker
source_timestamp	String	Original position in full ride (MM:SS-MM:SS)
duration_s	Float	Clip duration in seconds
telemetry_path	String	Path to telemetry CSV file
location	GeoLocations	GPS route polylines (front slice only)
event	TemporalDetections	Primary riding behavior
legality	TemporalDetections	Legal/illegal/unspecified
head_pose	TemporalDetections	Rider head direction
road_type	TemporalDetections	Paved/unpaved
road_marking	TemporalDetections	Marked/unmarked lanes
divider	TemporalDetections	Yes/no lane divider
traffic_density	TemporalDetections	Low/medium/high
n_lanes	TemporalDetections	Number of lanes

Frame-Level Fields

On front slice (telemetry):

• location: GPS coordinates (GeoLocation)
• speed_2d_mps: 2D speed in meters/second
• speed_3d_mps: 3D speed in meters/second
• gyro_x, gyro_y, gyro_z: Gyroscope readings
• telemetry_timestamp_s: Telemetry timestamp

On eye_tracker slice (gaze):

• gaze: Gaussian heatmap (224×224) showing attention region

Event Labels (12 Behaviors)

Conventional Behaviors (6):

• going straight: Normal forward motion
• left turn: Left turning maneuver
• right turn: Right turning maneuver
• lane change: Changing lanes
• u-turn: U-turn maneuver
• stop: Vehicle stopped

Unconventional Behaviors (6):

• overtake: Overtaking another vehicle
• weaving through traffic: Rapid lane switching in dense traffic
• obstruction avoidance: Swerving around obstacles
• violation: Traffic rule violation (e.g., signal jump)
• distracted driving: Rider distraction observed
• near collision: Close call with other road users

Legality Labels

• legal: Maneuver complies with traffic rules
• illegal: Explicit violation (weaving, signal jump, wrong lane)
• unspecified: Context-dependent (obstruction avoidance, near collision)

Note: This repository contains a subset of the full dataset. Use the import script to download and parse the complete dataset.

Setup Mapbox for GPS Visualization (Optional)

To visualize GPS routes in the FiftyOne App Map panel, you need a Mapbox API key:

1. Sign up for a free Mapbox account at https://mapbox.com
2. Get your API token from https://account.mapbox.com/access-tokens/
3. Export the token before launching the App:

export MAPBOX_TOKEN=your_mapbox_token_here

Dataset Creation

Curation Rationale

Two-wheelers account for a disproportionately high share of road fatalities in the Global South, yet research on rider behavior lags far behind four-wheeler ADAS research. The MOTOR dataset addresses this gap by providing the first large-scale, multi-view, multimodal resource for understanding two-wheeler behavior in dense, unstructured traffic conditions typical of countries like India and Indonesia.

The dataset enables research on:

• Understanding risky riding behaviors
• Predicting traffic violations
• Analyzing rider attention patterns
• Developing motorcycle-specific ADAS
• Improving road safety in emerging markets

Source Data

Data Collection

• Platform: Multi-camera setup with 3 GoPro Hero 10 cameras + eye-tracking glasses (Aria/Pupil)
• Duration: 4 weeks of data collection
• Location: Various roads in India (urban, highway, peak/off-peak hours)
• Riders: 16 riders with varying experience (2-20 years)
• Vehicles: Multiple two-wheeler types (motorcycles, scooters)
• Synchronization: All camera streams and sensors time-synchronized

Data Processing

• Video: 1920×1080, 30 FPS, with stabilization
• Telemetry: Extracted from GoPro recordings using GoPro telemetry extractor
• Gaze: Red gaze marker overlay burned into eye tracker video
• Clips: Pre-extracted to match event boundaries (typically 1-20 seconds)
• Annotations: Professional annotators trained on Indian Motor Vehicle Act (2017)

Annotations

Annotation Process

Two professional annotators labeled all sequences under expert supervision:

1. First 50 sequences annotated independently by both annotators
2. Expert reviewed work, resolved discrepancies, provided feedback
3. Remaining sequences annotated individually with periodic random checks
4. Annotations include:
   • Riding maneuver classification (12 classes)
   • Legality determination based on Indian Motor Vehicle Act
   • Head pose direction (on road, left, right, either side)
   • Traffic context (road type, lanes, markings, density)

Who are the Annotators?

Professional annotators trained on maneuver definitions and traffic violation rules from the Indian Motor Vehicle Act (2017), working under expert supervision.

Citation

BibTeX:

@inproceedings{paturkar2026motor,
  title={MOTOR: A Multimodal Dataset for Two-Wheeler Rider Behavior Understanding},
  author={Paturkar, Varun and others},
  booktitle={2026 IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026}
}

APA:

Paturkar, V., et al. (2026). MOTOR: A Multimodal Dataset for Two-Wheeler Rider Behavior Understanding. 2026 IEEE International Conference on Robotics and Automation (ICRA).