Skip to content

DATASETS.md

Latest commit

 

History

History
99 lines (81 loc) · 4.41 KB

DATASETS.md

File metadata and controls

99 lines (81 loc) · 4.41 KB
Raw

Datasets

All datasets share one OccStress root. Resource type comes first, then track, then dataset. Public dataset names are OccStress-nuScenes, OccStress-Waymo and OccStress-CARLA.

OccStress/
  meta/
    OccStress-nuScenes/{dataset.json,class_mapping.json}
    OccStress-Waymo/{dataset.json,class_mapping.json,controls.json}
    OccStress-CARLA/{dataset.json,class_mapping.json,controls.json}
  protocols/
    manual/<dataset>/<family>/<severity>/*.pkl
    upstream/<dataset>/<subtrack>/<source>/<condition>/*.pkl
    position_sweep/OccStress-nuScenes/<setting>/*.pkl
  occ/
    manual/<dataset>/<family>/<severity>/<scene>/<frame>/labels.npz
    upstream/<dataset>/<subtrack>/<source>/<condition>/<scene>/<frame>/labels.npz
  events/manual/<dataset>/...
  manifests/
    protocols.json
    assets.jsonl
    checksums/
    <dataset>.payload_progress.json
  external/<dataset>/...

Clean manual protocols use clean/H4_F6_val_backbone.pkl; traffic has no severity directory. Misalignment is represented by transforms, not duplicated occupancy. Upstream condition is clean or <corruption>/<severity>. Temporal patterns and position sweeps reuse assets; they do not add copies per model or protocol.

Mounts

From the checkout root:

export OCCSTRESS_CODE_ROOT="$PWD"
export OCCSTRESS_DATA_ROOT="/datasets/OccStress"

Alternatively, create data/OccStress as a local symlink only if that path does not already exist. Do not delete or overwrite an existing dataset to create it. The --occstress-root argument, where a native model exposes it, means this shared root, not a per-dataset directory.

Protocol paths are relative to this root, independent of the model process's working directory. Dataset names prevent collisions between identical scene or frame IDs. The resolver rejects mismatched dataset namespaces and parent traversal.

External Dependencies

Follow the step-by-step preparation commands. They cover safe mounts, raw-CARLA-to-canonical-GT conversion, and full clean-reference checks without changing released controls.

Original images, point clouds and official clean GT are not bundled in staging. Provide them under the following logical namespaces using authorized local mounts:

  • external/OccStress-nuScenes/gts/<scene>/<token>/labels.npz
  • external/OccStress-Waymo/native_gt/validation-data/...
  • external/OccStress-CARLA/canonical_gt/<scene>/<frame>/labels.npz
  • external/OccStress-CARLA/native_dataset/... for raw sensor dependencies.

OCCSTRESS_EXTERNAL_ROOT optionally relocates the external/ directory; it must contain the same OccStress-* subdirectories. nuScenes also requires official world-info metadata. Waymo/CARLA use meta/<dataset>/controls.json. Control arrays are preserved; loading JSON must not recompute poses, trajectories or commands. Canonical semantics is already Occ3D-18. Only native Waymo voxel_label needs mapping, exactly once. CARLA assets use the right-handed canonical view.

Validation and Availability

python tools/validate_occstress_dataset.py "$OCCSTRESS_DATA_ROOT"
python tools/validate_occstress_dataset.py "$OCCSTRESS_DATA_ROOT" \
  --dataset waymo --require-payloads

The default check reads the catalog and metadata, not all voxel files. During assembly, per-dataset progress and condition .done.json markers determine payload readiness. After finalization, manifests/status.json points to the consolidated finalization.json; an earlier stage=protocol_catalog snapshot is not a live completion signal. validation_coverage.json records sampled loader coverage separately. A completed payload is still not proof of full model-loader or GPU reproduction equivalence.

Finalized datasets include EXTERNAL_DATA.md and meta/external_dependencies.json. These describe authorized external mounts, not bundled originals or guaranteed availability on a new machine. CARLA command order is right/left/straight in the canonical x-forward/y-left frame; its control audit and any diagnostic metadata corrections are recorded under manifests/.

The current internal staging and sampled runtime checks are complete as recorded in release status. Published archives and their availability are listed in the download guide. Original clean GT remains an external dependency; publication does not replace source terms. Do not infer public availability from staging or model smoke tests.