The Indian Road Driving Dataset is the largest open dataset of annotated Indian road footage, created by ThirdEye Labs. It addresses the critical gap in autonomous driving datasets for Indian road conditions.

🌍 Why Indian Roads?

Indian roads present unique challenges absent from existing datasets (BDD100K, nuScenes, Waymo):

Dense mixed traffic with unpredictable behavior
Auto-rickshaws, cattle, and informal lane usage
Extreme lighting conditions
63 million vehicles and 1.4 billion people, yet no large-scale annotated dataset existed

📊 Dataset Statistics

Metric	Value
Total clips	8,437
Annotated frames	645,714
Object detections	3,669,119 (with ByteTrack IDs)
Segmentation masks	1,290,463
GPS-tagged frames	✅
Annotation format	BDD100K
Label provenance	Machine-generated, not human-verified
Sensor	Monocular camera + per-clip GPS (no LiDAR/radar/IMU)
Capture device	CP Plus dashcam
Location	Delhi NCR, India
Conditions	Day · Night · Dusk · Rain

⚠️ How these labels were produced

Labels are machine-generated and have not been human-verified. The export writes manualShape: false on every box, which BDD100K defines as machine annotation. Models used:

Stage	Model
Detection	YOLO11m fine-tuned on IDD
Tracking	ByteTrack
Segmentation	SegFormer-B5 fine-tuned on IDD
Scene attributes	CLIP ViT-L/14

Two consequences worth stating plainly:

Treat this as a large pseudo-labelled corpus, not gold ground truth. About 33% of detections fall below 0.5 confidence.
Do not benchmark detectors against these labels and call it accuracy. The labels come from IDD-fine-tuned models, so disagreement measures domain gap between label source and model under test, not correctness.

A human-verified evaluation split is in progress.

📈 Class coverage finding

Three of the 12 classes (autorickshaw, animal, vehicle fallback) have no equivalent in the BDD100K label set, so a detector trained on BDD100K cannot emit them at all.

Slice	Share of detections in unrepresentable classes
Overall	9.42%
Night	13.91%
Foggy	15.82%
Highway	11.10%
Clear daytime	8.57%

Auto-rickshaws alone account for 185,383 detections (5.05%) and appear in 6,775 of 8,437 clips (80.3%).

Counts computed directly from annotations/detection.json.

🏷️ Detection Classes (12 classes)

person: Pedestrians
rider: Motorcyclists/cyclists with rider
car: Passenger cars
truck: Trucks and tempos
bus: Buses
motorcycle: Motorcycles (unridden)
bicycle: Bicycles
autorickshaw: Auto-rickshaws (tuk-tuks)
animal: Cattle, dogs, animals on road
vehicle fallback: Unclassified vehicles
traffic light: Traffic signals
traffic sign: Road signs and boards

📁 Dataset Structure

Data is stored as 646 WebDataset tar shards (data/train-00000-of-00646.tar … data/train-00645-of-00646.tar), each containing ~1,000 frames. Each frame has 3 files inside the shard:

{clip_id}_{frame:04d}.jpg   # keyframe image
{clip_id}_{frame:04d}.png   # segmentation mask
{clip_id}_{frame:04d}.json  # BDD100K annotations (detections + scene attributes)

Standalone annotation files are also provided for convenient bulk access:

annotations/
├── detection.json          # BDD100K format, all 645,714 frames (1.2 GB)
└── scene_attributes.json   # per-clip weather, time of day, scene type
gps/
└── gps_tracks.json         # GPS coordinates per clip

🚀 Quick Start

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("thirdeyelabs/indian-road-dataset")
sample = ds["train"][0]
# sample keys: jpg, png, json

Load annotations directly

import json

with open("annotations/detection.json") as f:
    annotations = json.load(f)

# BDD100K format, each entry:
# { "name": "clip_id/frame", "labels": [{ "category": "car", "box2d": {...} }] }

Download with CLI

huggingface-cli download thirdeyelabs/indian-road-dataset --repo-type dataset

📐 Annotation Format (BDD100K Schema)

{
  "name": "clip_abc123/0042.jpg",
  "timestamp": 1000,
  "attributes": {
    "weather": "clear",
    "scene": "city street",
    "timeofday": "daytime"
  },
  "labels": [
    {
      "id": 1,
      "category": "car",
      "box2d": { "x1": 296.0, "y1": 242.0, "x2": 477.0, "y2": 379.0 },
      "attributes": { "occluded": false, "truncated": false },
      "track_id": 7
    }
  ]
}

🗺️ GPS Coverage

About two-thirds of clips include per-second GPS with speed and heading, enabling:

Geographic filtering by route/area
Speed and trajectory analysis
Map-based dataset exploration

🏗️ Production Pipeline

ThirdEye Labs end-to-end ML annotation system:

Ingest: raw MP4s from CP Plus dashcams to S3
Keyframe extraction: 1 frame/second via FFmpeg
GPS parsing: matched from .srt files
Object detection: custom YOLO fine-tuned for Indian roads
Semantic segmentation: SegFormer for drivable areas
Multi-object tracking: ByteTrack across frames
Scene classification: weather, lighting, scene type

📜 License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Free to use, share, and adapt for any purpose (including commercial) with attribution to ThirdEye Labs.

📚 Citation

@dataset{thirdeyelabs2026indianroad,
  title     = {Indian Road Driving Dataset},
  author    = {ThirdEye Labs},
  year      = {2026},
  url       = {https://huggingface.co/datasets/thirdeyelabs/indian-road-dataset},
  note      = {Released under CC BY 4.0}
}

🔗 Links

🌐 Website: thirdeyelabs.ai
🎬 Demo: thirdeyelabs.ai/demo
📧 Contact: thirdeyelabs.ai/contact

Built with ❤️ in India

Downloads last month: 439

Total file size:

210 GB