Skip to content

Releases

v0.1.2 alpha: recorded-camera timing and corpus compatibility

Choose a tag to compare

@gagandeepreehal gagandeepreehal released this 24 Oct 20:13
7f9d906

Camera exports now use the recorded-camera timestamp for every INDUS-AD recording, matching the original mz-datagen selection convention. Main GNSS retains its sensor timestamp. Untimed quality-flagged GNSS messages define excluded portions; usable portions remain eligible rather than rejecting the entire recording.

  • Preserve consistent nonzero GNSS origins, signed source offsets and missing diagnostics on quality-flagged messages without inventing data.
  • Keep exact source annotation identities while mapping verified source frames to recorded-camera export times.
  • Document timing, trimming, troubleshooting, manual access review and reproducible source compatibility results.

Validation: 1,039 tests, static analysis and offline export checks on Linux and macOS. Five representative real six-camera builds on macOS/Python 3.12 passed SDK loading, official nuScenes loading, all output-image decoding and CLI validation, yielding 229 samples and 1,374 images. Release evidence records the tested source revision and wheel hash.

This is an alpha prerelease. Metadata compatibility does not establish that every video has been decoded or that every recording yields an eligible scene. Production builds were tested with publisher credentials; no approved non-owner account was available for an access test. Hardware synchronization, physical calibration accuracy and privacy accuracy are not established by these software checks.

Python 3.12 is required. Dataset access is manually reviewed with no fixed turnaround. The software remains under PolyForm Noncommercial; the dataset license is unchanged. DOI registration remains on hold.

v0.1.1 — INDUS-AD devkit alpha

Pre-release

Choose a tag to compare

@gagandeepreehal gagandeepreehal released this 24 Oct 06:36
016ac6a

INDUS-AD devkit v0.1.1 is an alpha research release.

This release provides published calibration/GNSS adapters, a bounded first build, accurate Python support metadata, and Linux/macOS CI.

  • Quickstart: one six-camera recording, 181,048,771 unique selected artifact bytes (about 181 MB), with an empty human-label template and documented disk/runtime/output limits.
  • Python: 3.12 (>=3.12,<3.13). The official nuScenes loader check remains mandatory. Newer Python support is blocked by the official SDK dependency chain.
  • Dataset revision: ee0e1754964f6af000625fb42d854565b39fec36 in gagandeepreehal/indus-ad.
  • Release commit: 016ac6a9a38da70e1dc4d2251517f3af963f4551 (merged PR #13).
  • Attached distributions were built from 6eba9baabed536e052818d32706145897164c602; its complete tracked tree is identical to the release commit. SHA256SUMS records exact artifact hashes.

Validation: 1,028 local fresh-dependency tests; Ruff; strict mypy; generated schema; distribution builds; installed-wheel dependency check, offline export with official SDK loading, and CLI validation passed. PR CI and merged-main CI passed on Linux and macOS.

The quickstart selection was verified through metadata-only acquisition: one recording, six videos, 88 unique artifacts. No raw data, production videos or dense archives were downloaded. Runtime, peak disk use, and the selected recording's full production build and privacy quality remain unverified. The 2 GiB working budget is an estimate, not a guaranteed bound.

This software release does not constitute final dataset launch approval. The production withdrawal drill has not been executed; procedure preparation alone does not close that gate.