Metadata-Version: 2.5
Name: thirdeyelabs
Version: 0.1.2
Summary: Open road data for autonomous driving on unstructured Indian roads.
Project-URL: Homepage, https://thirdeyelabs.ai
Project-URL: Dataset, https://huggingface.co/datasets/thirdeyelabs/indian-road-dataset
Project-URL: Simulator, https://dream.thirdeyelabs.ai
Author: Third Eye Labs
License: Apache-2.0
Keywords: adas,autonomous-driving,computer-vision,dataset,india,object-detection
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Scientific/Engineering :: Image Recognition
Requires-Python: >=3.9
Provides-Extra: hf
Requires-Dist: datasets>=2.14.0; extra == 'hf'
Provides-Extra: sim
Requires-Dist: thirdeye-sim>=0.1; extra == 'sim'
Description-Content-Type: text/markdown

# thirdeyelabs

Open road data for autonomous driving on unstructured Indian roads.

Every major AV dataset (Waymo Open, nuScenes, Argoverse, PandaSet) was captured
on structured Western roads. This one was not.

```bash
pip install thirdeyelabs[hf]
```

```python
import thirdeyelabs as tel

tel.info()
# {'frames': 646014, 'clips': 8441, 'detections': 6896202, ...}

tel.INDIA_SPECIFIC
# ['autorickshaw', 'animal', 'vehicle_fallback']

ds = tel.load_dataset()
```

## The dataset

| | |
|---|---|
| Frames | 645,714 across 8,437 clips |
| Detections | 3,669,119 2D boxes with track IDs |
| Segmentation | 1,290,463 masks |
| Classes | 12 detection, 36 segmentation. BDD100K schema, extended for Indian roads |
| Region | Delhi NCR, India |
| Sensor | Monocular dashcam, GPS per clip |
| License | CC BY 4.0 |

Three of the twelve categories (`autorickshaw`, `animal`, `vehicle fallback`)
have no equivalent in BDD100K, so a detector trained on it cannot emit them at
all. They account for **9.4% of all detected road agents**, rising to 13.9% at
night and 15.8% in fog. Auto-rickshaws alone appear in 80% of clips.

Labels are machine-generated by models fine-tuned on Indian road data, not
human-verified.

Dataset: https://huggingface.co/datasets/thirdeyelabs/indian-road-dataset

## Dream, the driving simulator

Dream is a Gymnasium driving environment on Gurgaon's real roads, in traffic that
behaves like India's, with event rates measured from this footage. Play it in
your browser at https://dream.thirdeyelabs.ai, or install it:

```bash
pip install "thirdeyelabs[sim]"
```

```python
import gymnasium as gym
import thirdeye_sim.env

env = gym.make("ThirdEye/Gurgaon-v0")
obs, info = env.reset(seed=0)
```

Code and documentation: https://github.com/thirdeyelabsai/thirdeye-sim

## Status

Early. This release provides dataset access helpers, the taxonomy and the
driving simulator (through the `sim` extra). Benchmark suites are in
development.

## License

Package: Apache-2.0. Dataset: CC BY 4.0.
