| --- |
| pretty_name: TUV stationary-start vehicle test clips |
| language: |
| - en |
| tags: |
| - autonomous-driving |
| - radar |
| - neural-reconstruction |
| configs: |
| - config_name: clips |
| default: true |
| data_files: |
| - split: test |
| path: clips.csv |
| --- |
| # Stationary-start TUV test clips |
|
|
| This dataset contains 86 clips with verified zero-speed starts and requested event durations below 35 seconds, plus one explicitly selected 35-second CCRb clip, across CCRs (stationary target), CCRm (moving target), and CCRb (braking target). |
|
|
| CCRs: 31, CCRm: 50, CCRb: 5 |
|
|
| ## Start here |
|
|
| Read `clips.csv` to select a clip by scenario, speed, or duration. All paths in the manifest are relative to this dataset root. `files.csv` lists each payload file and byte size. The `test` split is an organizational choice; no train/test independence claim is made. |
|
|
| ```text |
| data/<scenario>/<clip_id>/ |
| nurec/ |
| last.usdz Original complete NuRec bundle |
| rclog/ |
| roadcast_debug.log Recorded RC log for the requested event window |
| radar/ |
| radar_front_center.bin |
| radar_corner_front_left.bin |
| radar_corner_front_right.bin |
| radar_corner_rear_left.bin |
| radar_corner_rear_right.bin |
| *.bin.seek Available radar seek indexes |
| ground_truth/ |
| <test>.csv Original full GT measurement CSV |
| <test>_window.csv Available GT samples within the estimated event window |
| alignment.json Clock conversion and GT coverage gaps |
| ``` |
|
|
| The folder name is the clip UUID, with `--IVEX` appended for shortlisted clips; the source event and recording session UUIDs are separate columns in `clips.csv`. Start/end values are PTP microseconds. Duration is `(end_ptp_us - start_ptp_us) / 1e6`. Sensor extraction can include format/index boundary data; use timestamps and seek indexes when reading. |
|
|
| `last.usdz` contains the complete packaged reconstruction. The USD scene, checkpoint, trajectory/tracks, configuration, and other available assets remain inside the USDZ. Extract the bundle locally only if your tools require individual files. Not every reconstruction has identical optional assets. A compatible NuRec renderer is needed for neural rendering; a generic USD viewer may not render the neural appearance. Radar and RC logs require compatible decoding tools and are not CSV tables. |
|
|
| ## Ground truth and limitations |
|
|
| GT CSVs retain their original columns and units (the second CSV row contains units). `_window.csv` preserves only available recorded samples within an estimated clock-aligned interval; it is not extrapolated to fill the clip. Alignment uses speed traces, not an exact hardware clock calibration. Read `alignment.json` and manifest `gt_missing_start_s` / `gt_missing_end_s` before comparing trajectories. Having a matched GT CSV does not guarantee full stationary-window coverage or prove trajectory agreement. |
|
|
| Stationary refers to ego speed at the requested event start. It does not mean the entire scene or target remains stationary. The reconstructed camera sequence can have a slightly narrower frame-available window than the requested recording interval. |
|
|
| Inclusion requires NuRec, recorded RC log, all five radar bins, with GT where a confident match exists. Five selected clips have missing GT; these are flagged by `gt_status=missing`, an empty `gt_window_path`, and `ground_truth/MISSING_GT.txt`. Do not treat them as ground-truth-evaluated clips. Clips missing NuRec or verified recordings are excluded until those artifacts are available. This selection is not a claim of KPI success or visual quality. No PSNR values are supplied; the upstream PSNR computation was unavailable. Perception-agreement failures do not establish missing rendered obstacles. |
|
|
| ## Load the index |
|
|
| ```python |
| from datasets import load_dataset |
| clips = load_dataset("YOUR_ORG/YOUR_DATASET", "clips", split="test") |
| print(clips[0]) |
| ``` |
|
|
| Download only the files needed for one clip: |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| root = snapshot_download( |
| repo_id="YOUR_ORG/YOUR_DATASET", |
| repo_type="dataset", |
| allow_patterns=["clips.csv", "data/CCRs/CLIP_UUID/**"], |
| ) |
| ``` |
|
|
| Upload this directory as individual files, retaining its structure. It is not necessary to ZIP the dataset. Licensing and distribution terms must be supplied by the dataset publisher; this card does not assign a license. |
|
|
| ## IVEX shortlist |
|
|
| Folders ending in `--IVEX` identify the 15 executive-review selections. The additional CCRb clip `8fc7e6c9-e4df-41a8-b11f-c2d650c4f53e` is a verified 35-second exception. |
|
|
| The 15 clips marked `--IVEX` use NuRec `26.8.172-6ada330c-lidarfree`. Their requested recording windows and companion recordings and GT files are unchanged. |
|
|
| An additional 50 non-IVEX clips (26 CCRs and 24 CCRm) use NuRec `26.8.172-6ada330c-lidarfree`. Requested recording windows, RC logs, radar data and GT files are unchanged. |
|
|
| The remaining 21 CCRm reconstructions have also been updated. All 86 clips now use NuRec `26.8.172-6ada330c-lidarfree`; requested windows and companion data are unchanged. |
|
|
