This repository contains nuImages source images and derived corruption images. Access and use remain subject to the controlling nuImages Dataset Terms. The dataset is for non-commercial research use; this repository does not grant a broader license or imply endorsement by Motional.

Continuous Condition-Severity nuImages Corruption Grid

This public, gated dataset contains the frozen image and metadata release used for continuous condition/severity adapter experiments with PicoDet-M. Access is automatically granted after a signed-in user acknowledges the terms above.

Scope

Clean nuImages source images used by the frozen training and development split.
Six derived corruption families: ColorQuant, Fog, LowLight, MotionBlur, Snow, and Brightness.
Ten normalized severity levels per corruption in the complete evaluation grid.
COCO-format object-detection annotations, frozen manifests, configurations, and SHA-256 inventories required to reproduce the experiments.
Approximately 430.4 GB in 26 portable tar shards. The three subset_* shards and 23 full_* shards are disjoint by construction.

These are portable tar archives rather than a normalized WebDataset key/value layout. The tar members retain their release-relative carla_data/... paths. Files in manifests/ list every member and its original SHA-256. SOURCE_COMPLETE.json records source counts, split checks, byte counts, and manifest identities.

This is a development dataset. It is not an untouched test or qualification set. The frozen adapter training view contains 12,264 source images and 85,848 views. The matched evaluation contains 650 source images across 61 cells: Clean plus six corruptions at ten severity levels.

Quick start: reproduce the Balanced7 training and evaluation inputs

The frozen Balanced7 input is already included in this repository. It is not a separate dataset repository or a random sample that needs to be regenerated. For the current PicoDet-S adapter experiments, download all three subset archives (the selected images are split across them):

hf auth login
hf download apdoa/continuous-cg-full-grid --repo-type dataset \
  --include 'data/subset_chunk_*.tar' \
  --include 'manifests/subset_chunk_*.files' \
  --include 'manifests/subset_chunk_*.sha256' \
  --local-dir continuous-cg-full-grid

mkdir -p continuous-cg-extracted
for archive in continuous-cg-full-grid/data/subset_chunk_*.tar; do
  tar -xf "$archive" -C continuous-cg-extracted
done

The extracted tree contains carla_data/... paths. The canonical training annotation is:

<extract-root>/carla_data/
  nuimages_robobev_continuous_g1_v1_20260903_experiments/
    balanced7_adapter_comparison_v2_20260904/
      data/seed_7301_balanced7/coco/train/balanced7_universal.coco.json

It contains 85,848 training images/views. The matched evaluation annotations are under evaluation_protocol/calibration/coco/: one Clean cell and six conditions at ten severity levels (61 cells total, 39,650 images/views). The selected training images occupy about 31.5 GiB and the matched evaluation images about 15.5 GiB; reserve at least 60 GiB after extraction for a complete local reproduction plus working space.

The COCO annotations preserve the source machine's absolute file_name prefix. Keep the archive JSON unchanged for checksum provenance. If extracting under a different root, create a local copy of each annotation JSON and rewrite only that absolute prefix to <extract-root>/carla_data/...; do not change image IDs, annotations, condition/severity labels, or ordering. Point the training configuration at the rewritten Balanced7 copy and use the equivalent rewritten evaluation JSONs. Changing only anno_path does not rewrite image paths.

full_* archives are unnecessary for this Balanced7 reproduction; they are for reconstructing the rest of the complete corruption grid. Verify the relevant archive/member manifests before training.

License and redistribution notice

The underlying nuImages data are provided by Motional under the controlling nuImages Dataset Terms. The official nuImages page describes the dataset as strictly non-commercial, and the terms state that derived data remain subject to those terms. This repository does not relicense the nuImages images or derived images and does not grant commercial-use rights. Users are responsible for reviewing the current controlling terms, preserving required attribution, checking any third-party exceptions in the source metadata, and obtaining a commercial license from Motional when applicable.

Motional/nuImages does not sponsor, approve, or endorse this release. For license questions, use the contact information provided on the official nuImages website.

Citation

Please cite nuImages/nuScenes as requested by the official dataset documentation and cite the associated adapter work when its paper citation becomes available.

Limitations

Images can contain people, license plates, locations, and other information present in the original driving dataset.
Corruptions are synthetic transformations and do not cover all real adverse conditions.
Severity numbers are condition-specific controls; equal numbers across two conditions do not necessarily represent equal physical degradation.
The development cells reuse source scenes and are not independent samples for statistical significance testing.

Temporal Corruption Episodes (TCE v1)

This release adds a frozen, evaluation-only temporal input package under tce/v1_20260920/. It contains 40,769 native nuImages CAM_FRONT source images, the calibration/reporting episode metadata, and the SHA inventories used for the Temporal Corruption Episodes (TCE) switching-policy evaluation. It is not a new training split and it does not contain a sealed test set.

The archive is split into 1-GiB pieces for resumable download. The pieces must be concatenated in lexical order before extraction. The release manifest states the expected whole-archive SHA-256, image inventory SHA-256, split roles, and the per-piece hashes.

hf download apdoa/continuous-cg-full-grid --repo-type dataset \
  --include 'tce/v1_20260920/**' --local-dir continuous-cg-full-grid

cd continuous-cg-full-grid/tce/v1_20260920
sha256sum -c manifests/tce_v1_20260920.chunks.sha256
cat data/tce_v1_20260920.tar.gz.part-* > tce_v1_20260920.tar.gz
sha256sum -c manifests/tce_v1_20260920.archive.sha256
tar -xzf tce_v1_20260920.tar.gz

The TCE images and derived metadata remain subject to the same nuImages terms, non-commercial research restriction, and gated-access acknowledgement stated above. The temporal recipes are deterministic evaluation artifacts: reporting is for frozen-policy evaluation, while calibration is for policy selection only.

Downloads last month: 28

Total file size:

439 GB

Models trained or fine-tuned on apdoa/continuous-cg-full-grid

apdoa/picodet-m-aj-epoch5

Object Detection • Updated Sep 6