Skip to content

data.md

Latest commit

 

History

History
53 lines (43 loc) · 2.21 KB

data.md

File metadata and controls

53 lines (43 loc) · 2.21 KB
Raw

Data preparation

The release contains the benchmark annotations, schema, source-video references, privacy masks and preprocessing utilities. Download the fixed-version archive using bash scripts/prepare_data.sh; the downloader verifies its SHA-256 and installs all ten JSONL files under annotations/.

Split Train Test
L1 Core 4,843 1,024
L2 Core 5,211 1,073
L3 Core 5,003 1,949
L4 Core 5,003 1,949
L3 OOD — 399
L4 OOD — 399

Each record contains question_id, question_level, video_name, question_type, split, sample_type, alert_mode, source_type, image_samples, t_ref_frame, question, answer and trigger_candidate_count. image_samples contains ordered paths relative to --dataroot; t_ref_frame indexes the source frame sequence at 10 FPS.

Obtain Core camera images from nuPlan. Preserve each relative image path under the selected dataroot. OOD source links, frame extraction and privacy-mosaic instructions are installed under external/ood/:

OOD preprocessing uses Python 3.10 or later, FFmpeg on PATH, NumPy and OpenCV. Install the Python dependencies with python -m pip install numpy opencv-python-headless.

python external/ood/code/extract_resize_682x384.py \
  --input-path /path/to/source_videos --output-path /path/to/frames/bilibili \
  --annotation-jsonl annotations/l4_ood.jsonl
python external/ood/code/mask_privacy.py \
  --mask-dir external/ood/file/masks --img-root /path/to/frames \
  --output /path/to/redacted_frames

See external/ood/OOD_data_process.md for the source-video and frame-layout details.

python -m drivestreambench.data.validate annotations/l4_core_test.jsonl
bash scripts/build_l4_short.sh

Short positives retain a window around the reference frame; short negatives retain their first 15 seconds. metadata.crop records the window and remapped reference index. Use DSB_SHORT_SECONDS to select another duration.

The data archive includes FILES.sha256 and a split index. Croissant metadata is a separate release asset. The benchmark-created annotations and utilities are CC BY 4.0; source-media rights remain with the original providers.