gagandeepreehal/indus-ad-sample

INDUS-AD Sample

Try the INDUS-AD multi-camera workflow without requesting access. This public sample contains eight complete recordings from Bengaluru, with camera videos, timing and calibration tables, privacy-transformed GNSS/pose data, sampled pseudo-labels, and their semantic/depth maps.

Request access to the full INDUS-AD dataset →

The sample downloads independently. It requires no approval or access to the main repository. Full-dataset access is reviewed separately.

What is included

Content	Sample
Complete recordings	8
Recording timeline	16 min 43.7 s
Camera MP4s	55
Camera coverage	Six surround cameras in every recording; tilted front in seven
Sampled camera annotation rows	140
Semantic masks / depth maps	140 / 140
GNSS messages, including both published channels	60,224
Download size	9.53 GB (decimal)

Recording time counts each recording once, rather than summing camera views. See the exact file manifest for byte sizes, SHA-256 hashes, recording IDs, selection method and source revision.

The recordings were chosen for manageable size, camera/annotation coverage and a range of daytime, evening and night material. Source weather labels helped screen candidates and remain unvalidated model predictions. This is a curated workflow sample, not a statistically representative subset or an authored benchmark split. The Viewer's technical train split and the source sample_group folders do not define evaluation partitions.

Start with one camera

Install Python dependencies and have ffmpeg available on your PATH:

pip install huggingface_hub pyarrow pillow

Download and run the example without logging in:

import runpy
from huggingface_hub import hf_hub_download

script = hf_hub_download(
    "gagandeepreehal/indus-ad-sample", "examples/quickstart.py",
    repo_type="dataset", token=False,
)
runpy.run_path(script, run_name="__main__")

The script downloads one front-camera MP4 and the matching GNSS/annotation tables, reports coverage and saves sample-preview/preview.jpg. It resolves the sample revision once for consistent downloads. To preview all six surround cameras, download quickstart.py and run python quickstart.py --surround.

For a complete local copy:

from huggingface_hub import snapshot_download
snapshot_download(
    "gagandeepreehal/indus-ad-sample", repo_type="dataset",
    local_dir="indus-ad-sample", token=False,
)

Structure and compatibility

The main dataset's directory layout, recording IDs, camera names, source table schemas, annotation frame IDs and timestamp values are preserved. All selected videos are byte-identical copies of the published main-dataset videos. Dense archives are repacked to contain only selected members; the member bytes and original payload hashes are preserved, and archive references are updated. Viewer controls and manifests describe this sample, and video references resolve inside this repository.

data/metadata.jsonl maps recordings to video paths relative to data/. Normalized tables are under data/tables/; sampled labels, dense indexes and TARs are under data/annotations/camera/. Viewer configurations are videos, camera_annotations, gnss and semantic_depth.

Use the INDUS-AD devkit with this repository ID and a sample recording ID. The layout/control contract is checked separately from a complete nuScenes export; no successful full export is claimed for every sample recording. The standalone quickstart does not require the devkit.

Camera alignment uses rec_time_offset_ns; the main gnss channel uses sensor time_offset_ns. Do not substitute logger time or combine rec_gnss with the main trajectory. The flattened GNSS Viewer preserves the main dataset's schema and omits channel; use the per-recording source tables to distinguish channels. Timing policy · Sensor conventions.

Limitations, licence and attribution

Annotations are sparse machine-generated pseudo-labels, not human ground truth. Unlabelled frames and cameras are expected. Weather scores are not validated probabilities; predicted depth is not measured depth ground truth. No steering, throttle or brake targets are included.

The videos retain the source's automated face/plate blur. Blur can miss identifying content, and published trajectories retain linkage risk. Sampled visual inspection and integrity checks do not certify complete anonymisation or physical calibration accuracy. This sample is for non-commercial research, teaching and evaluation of the data workflow; it is not validated for production driving or safety decisions.

Released under CC BY-NC 4.0. Attribute Gagandeep Reehal, INDUS-AD (2026) and link to the full dataset. This release selects records, repacks dense archives and rebuilds sample controls; it does not alter video or dense-member payloads. Third-party annotation notices remain applicable.

Do not re-identify people, track individuals or expose sensitive locations. Report privacy/takedown concerns privately to indus-ad@gagandeepreehal.com, including the sample repository and recording ID; do not post identifying evidence publicly. Privacy and takedown policy.

Ready for more recordings? Request full INDUS-AD access.