INDUS-AD: Indian Dataset of Unstructured Urban Scenes for Autonomous Driving

5.5K+ total downloads as of October 2026.

Overview

INDUS-AD—Indian Dataset of Unstructured Urban Scenes for Autonomous Driving—is a 4.006 TB research dataset built from 1,469 driving recordings across Bengaluru (Bangalore), India. It brings together 9,927 face/plate-blurred multi-camera MP4s, 588.57 aggregate camera-view hours, normalized sensor tables, and derived semantic and depth layers in one directly usable release.

Each recording is organized around a stable recording_id that joins camera video with camera timing and calibration, privacy-transformed GNSS/pose fields, frame-level camera pseudo-labels, and dense semantic/depth archives. The result supports non-commercial research in end-to-end driving, video and world models, multi-camera perception, localization, synchronization, and robotics data tooling.

INDUS-AD is research data rather than a ready-made benchmark. It does not prescribe train/validation/test partitions or provide documented synchronized steering, throttle, and brake targets. Researchers retain control over filtering, sampling, synchronization tolerances, quality rules, target construction, and splits. Camera annotations, semantic outputs, depth outputs, and weather fields are model-generated weak pseudo-labels—not human ground truth.

Example six-camera surround view from Bengaluru. The image illustrates the urban traffic context and camera perspectives; it is not an authored benchmark label or a claim of uniform geographic coverage.

At immutable video-content revision 4ef41cd594234980f416136a543e3374be01735a, all 9,927 published MP4s are face/plate blurred. Blur regions are model-generated and can miss or over-blur content; gated access remains an access control, not a claim that visible or geospatial privacy risk has been eliminated.

Key facts

Property	Value
Recordings / video Viewer rows	1,469
Camera MP4 files	9,927
MP4 payload bytes	3,956,842,789,114
Approximate published repository size	4.006 TB
Recording timeline (longest camera stream per recording)	85:18:25.609 (85.31 hours)
Aggregate camera-view duration	588:34:01.225 (588.57 hours)
Face/plate-blurred MP4 files	9,927
Named camera positions	7
GNSS recording tables	1,469
Camera-frame tables	9,927
Frame-label tables	1,469
Camera-annotation Viewer rows	34,306
Semantic/depth Viewer rows	34,306
GNSS Viewer rows	18,373,026
Dense-index tables	69
Dense depth archives	68
Dense semantic archives	1
Authored benchmark split	None

Ways to explore INDUS-AD

INDUS-AD is an MP4-and-table dataset. Researchers can start in the Viewer, download selected camera videos, work directly with normalized sensor tables, or resolve full semantic and depth members through the dense indexes.

Dataset resources

The following supporting resources are packaged with this dataset:

• camera-duration evidence;
• recording-level quality summary and its integrity manifest;
• sensor and timing conventions;
• privacy and recording-level takedown policy; and
• release receipt, which binds the control-plane files to the immutable video tree.

Camera coverage

Conceptual camera layout, not calibrated geometry or a to-scale field-of-view diagram. INDUS-AD additionally includes cam_front_tilted for 1,113 recordings.

Camera	Recordings with video	Missing recordings
cam_front	1,469	0
cam_front_left	1,469	0
cam_front_right	1,469	0
cam_front_tilted	1,113	356
cam_rear	1,469	0
cam_rear_left	1,469	0
cam_rear_right	1,469	0

Missing camera data is represented as null. Another camera is never copied or fabricated to fill an absent camera position.

Bengaluru coverage

The recordings span urban roads across Bengaluru (Bangalore). Brighter areas indicate greater concentration in this aggregate visualization; the map does not imply uniform sampling, exhaustive city coverage, or a calibrated density scale.

Driving-condition examples

These paper-selected frames illustrate visual variation in the release. The condition names come from model-generated camera pseudo-labels, not human-authored weather ground truth, and the panel is not a corpus-prevalence estimate.

Researcher-controlled data policy

INDUS-AD publishes reusable layers rather than prescribing one benchmark pipeline. Document sensor selection, temporal windows, synchronization tolerance, quality rules, target construction, and split policy for every derived task dataset.

Dataset structure

data/
  metadata.jsonl
  manifests/
    video-integrity.jsonl
    video-privacy-release.json
    quality-summary/
      camera_duration_summary.csv
      quality_summary_manifest.json
  viewer/
    camera-annotations/*.parquet
    gnss/*.parquet
    semantic-depth/*.parquet
    videos/*.parquet
  videos/
    <camera>/<two-hex-bucket>/<recording_id>.mp4
  tables/
    auxiliary-messages/
    calibrations/
    camera-frames/
    camera-messages/
    gap-events/
    gnss/
    keyframes/
    recording-metadata/
    recording-quality-summary/
    sync-groups/
  annotations/camera/
    frame-labels/
    dense-index/
    depth/
    semantic/

The date-free recording_id is the stable join key across video, GNSS, camera timing, frame labels, and dense archive indexes. The two-character video bucket is sha256(recording_id)[:2]; it is only a deterministic storage partition. Use data/metadata.jsonl rather than guessing paths.

No benchmark result tables or authored benchmark partitions are included. Any split value inside imported camera pseudo-label records is a deterministic storage/sampling label from the source annotation process, not an official performance benchmark split.

Dataset Viewer

The Viewer has four separate configurations so rows with different meanings are never flattened into one table:

• videos (default): one row per recording with recording_id and seven camera video columns;
• camera_annotations: one row per sampled camera frame with semantic, depth, weather, and image-quality pseudo-label fields;
• gnss: one row per GNSS message with timing, privacy-transformed coordinates, motion, orientation, validity, and RTK/accuracy fields;
• semantic_depth: one row per sampled camera frame with a colorized semantic-mask preview, a colorized metric-depth preview, exact join keys, and the per-frame depth range in meters.

Viewer tables intentionally omit SHA-256 values, checksums, byte sizes, schema hashes, model/checkpoint hashes, archive internals, and processing diagnostics. The original integrity metadata remains available at data/manifests/video-integrity.jsonl, but it is outside every Viewer configuration.

Missing tilted-camera values remain null. Video paths in data/metadata.jsonl are relative to data/.

Download one camera video

import json
from huggingface_hub import HfApi, hf_hub_download

repo_id = "gagandeepreehal/indus-ad"
revision = HfApi().repo_info(
    repo_id,
    repo_type="dataset",
    revision="main",
).sha
recording_id = "17-53-27.393-d655e06f8e7ee863e68443502137f2fe"

metadata_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    revision=revision,
    filename="data/metadata.jsonl",
)

with open(metadata_path, encoding="utf-8") as stream:
    rows = (json.loads(line) for line in stream)
    row = next(row for row in rows if row["recording_id"] == recording_id)

relative_path = row["cam_front_video_file_name"]
video_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    revision=revision,
    filename=f"data/{relative_path}",
)
print(video_path)

The example requires accepted gated access and a logged-in Hugging Face account. It resolves main once, then uses that immutable commit for both downloads. Replace recording_id with any value from data/metadata.jsonl. Download narrow file patterns rather than cloning the complete multi-terabyte repository.

Timing, joins, and GNSS privacy transformation

Calendar dates were removed from published recording identifiers and repository paths. Time remains intact: tables retain time-of-day nanoseconds and per-recording nanosecond offsets so camera frames, GNSS messages, and pseudo-label rows can be aligned without reconstructing a capture date.

Across the 9,927 recording-camera streams, the per-camera median cadence is approximately 10.0007–10.0008 FPS. Across 3,028,347 decoded producer-batch groups, the median timestamp spread is 0.020 ms and the p95 spread is 0.025 ms. These are descriptive timestamp measurements—not evidence of a shared hardware trigger, exposure simultaneity, calibrated clocks, or end-to-end latency.

GNSS publication follows these rules:

• precise local ENU coordinates remain fully precise;
• global ENU and global latitude/longitude are published at approximately 1 m horizontal resolution;
• height is retained;
• sensor-quality and RTK diagnostics such as east/north sigma, position RMS, HDOP, and fix type are retained;
• published_horizontal_resolution_m records the 1.0 m publication resolution;
• rows with missing source coordinate objects remain null and invalid—coordinates are never invented; is_valid=false can also occur on rows that still contain sensor-reported coordinates.

Preserve camera identity and published timestamps when joining. Do not synchronize by filename order, protobuf array position, decoded-frame index, or an assumed constant frame rate. Cameras can begin late, end early, temporarily disappear, or contain timestamp gaps.

Camera pseudo-labels and dense archives

The frame-label tables contain model-generated weak pseudo-labels and summary features, not human ground truth. They include semantic class fractions, metric-depth summaries, weather probabilities, image-quality values, timing/join keys, and archive members. The dense index maps frame rows to full-resolution semantic or depth members stored in TAR archives.

The semantic_depth Viewer configuration provides 320×180 display previews for all 34,306 paired semantic masks and depth maps. The semantic_mask image column preserves class IDs as palette indices. The depth_map image column visualizes the normalized 16-bit map, while depth_min_m and depth_max_m give its corresponding metric range. These previews are for browsing—the authoritative 1280×720 uint8 semantic masks and uint16 depth maps remain in the dense TAR archives.

Model provenance identifies which model and immutable model revision produced each derived field:

• semantic: facebook/mask2former-swin-large-mapillary-vistas-semantic at revision 4772b6bf101d91f2534c106dc524d906aeb3c68a;
• depth: depth-anything/Depth-Anything-V2-Metric-VKITTI-Large at revision 070e97e4b80e317ec1d03c19927304f4091a180b;
• weather: openai/clip-vit-base-patch32 at revision 3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268.

These outputs can contain model errors, dataset biases, calibration limitations, and missing joins. Validate them for the intended research use and do not describe them as manually verified labels.

Intended uses

Appropriate non-commercial uses include end-to-end trajectory research with carefully derived pose-based future-motion targets; video foundation models and world models; camera perception and representation learning; temporal and cross-camera synchronization; GNSS/pose alignment; localization experiments; calibration; data-quality tooling; teaching; and research demonstrations.

The dataset is not validated for production driving systems, operational safety decisions, surveillance, identity recognition, or unsupported performance claims.

Quality, sensor, and takedown references

• recording_quality_summary.parquet contains date-free per-recording duration, camera completeness, gap, GNSS-validity, and decoded synchronization summaries for all 1,469 recordings.
• camera_duration_summary.csv distinguishes per-camera duration, the recording timeline, and aggregate camera-view duration.
• SENSOR_AND_TIMING_CONVENTIONS.md documents camera identity, synchronization evidence, GNSS publication, calibration boundaries, and unknown hardware metadata.
• DATASET_PRIVACY_AND_TAKEDOWN.md defines private reporting, recording-level withdrawal, cross-release removal, verification, and denylisting against reintroduction.

The recording-quality summary is an aggregate diagnostic table, not a benchmark result table or authored train/test split.

Limitations and responsible use

• Geographic, environmental, temporal, traffic, route, and hardware coverage may be uneven.
• Frames or table rows can be missing, delayed, duplicated, corrupted, or irregularly spaced.
• Calibration, sensor availability, and decoded schemas can differ between recordings.
• Videos were processed by automated face/plate detectors but can retain missed identifying content or contain false-positive blur.
• Reduced-resolution global coordinates and gated access reduce some exposure but do not make trajectories anonymous; local ENU and external data can still create linkage risk.
• Column presence does not guarantee a populated value in every row.

Do not use the dataset for re-identification, tracking individuals, exposing sensitive locations, or combining it with external data in ways that increase privacy risk. Users are responsible for applicable privacy, data-protection, publicity, export-control, and local recording laws.

Report sensitive or improperly included material privately through the Minus Zero contact form or team@minuszero.ai with subject Dataset privacy or takedown request. Include the repository and date-free recording_id or path when known. Do not post identifying evidence in a public discussion.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). Attribution is required, commercial use is prohibited, and modifications must be indicated. The license does not remove privacy, publicity, trademark, or other third-party rights.

Citation

@dataset{indus_ad_2026,
  author    = {Gagandeep Reehal},
  title     = {INDUS-AD: Indian Dataset of Unstructured Urban Scenes for Autonomous Driving},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/gagandeepreehal/indus-ad}
}

Maintenance

Use the repository discussions for non-sensitive questions and corrections. Use the private contact route above for privacy and takedown reports.