xlu11/carla_collision

CARLA Collision Video, Pose Pairs and Scene Captions

2,248 source videos and 3,764 captioned five-second windows, collected in CARLA 0.9.16 at 16 Hz, 1280 x 704, 90-degree horizontal FOV. Each source contains a deliberately constructed collision encounter, actual and constructed input camera trajectories, physical vehicle controls, and collision sensor events. Each indexed window has an English scene_static caption generated from all 81 consecutive images by Qwen/Qwen3.8-27B.

Contents

Path	Contents
data/*.tar	32 original, uncompressed archives; 31.43 GB (29.27 GiB)
dataset.json	Original archive manifest, counts, sizes and SHA-256 hashes
samples.jsonl	2,248 source records, archive locations and per-file hashes
clips_5s.json	3,764 windows, collision labels, splits and 93 excluded sources
captions/scene_static_81.jsonl	Complete captions joined to the current window index
tables/train.parquet	3,623 training rows with the same fields as the caption JSONL
tables/validation.parquet	141 validation rows with the same fields
captions/provenance.jsonl	Original model responses, evidence and input PNG hashes
captions/annotation_recipe.json	Exact prompts, generation settings and production hashes
tools/	Portable annotation client, direct archive reader, verifier and tests
SHA256SUMS	Integrity manifest for every release file except itself
release.json	Release counts, licensing and source artifact provenance

These are ordinary tar archives containing samples/<map>/<sample_id>/<file>; they are not WebDataset shards. Videos are preserved byte for byte. The five-second windows are indexes into source MP4 files, not re-encoded video copies. Temporary annotation PNGs and extracted duplicates are omitted.

Splits

Split	Contributing source videos	Windows	First-collision windows	Pre-collision windows
Train	2,073	3,623	2,073	1,550
Validation: Town11 and Town15	82	141	82	59
Total	2,155	3,764	2,155	1,609

All windows from a source stay in one split. Town11 and Town15 are held out; the other eight maps are training only. There is no independent test split. Index and JSONL rows use split="val"; the Hugging Face split is named validation. These refer to exactly the same 141 rows.

The original caption export preceded the split assignment. This release updates only its split field from the current index. Every caption string, frame boundary and original model response is preserved. Original and release index hashes are recorded separately in captions/annotation_recipe.json.

Load the Tables

After publication, use the actual repository ID in place of OWNER/DATASET:

from datasets import load_dataset

dataset = load_dataset("OWNER/DATASET")
row = dataset["train"][0]
print(row["clip_id"], row["scene_static"])

The tables expose caption text and metadata in the Hub viewer. video_path is a relative path string, not an embedded video column. To decode videos or read poses, also download the corresponding tar archive or the full snapshot.

For an already downloaded release:

from datasets import load_dataset

dataset = load_dataset("parquet", data_files={
    "train": "tables/train.parquet",
    "validation": "tables/validation.parquet",
})

Download and Read Video/Pose Data

from huggingface_hub import snapshot_download

snapshot_download("OWNER/DATASET", repo_type="dataset", local_dir="carla-collision-pose")

From the downloaded directory, with Python 3.11 or newer:

python -m pip install -r requirements.txt
python tools/verify_release.py
python tools/read_clip.py --split train --index 0

read_clip.py reads one source directly from its archive when it has not been extracted, decodes the exact 81 images, and reads its 81 states and 80 actions. The Python function returns RGB images and both actual and input c2w trajectories expressed relative to the respective first pose, without changing metric scale:

from tools.read_clip import read_clip

sample = read_clip(".", split="train", index=0)
print(sample["rgb"].shape)                 # (81, 704, 1280, 3)
print(sample["input_c2w"].shape)           # (81, 4, 4)
print(sample["actual_c2w"].shape)          # (81, 4, 4)
print(len(sample["actions"]))             # 80
print(sample["caption"])

For repeated training or re-annotation, extract once:

mkdir -p extracted
for shard in data/*.tar; do tar -xf "$shard" -C extracted; done

All index paths, including extracted/, are relative to the directory containing clips_5s.json. Full extraction requires approximately another 29.26 GiB.

Source Sample Format

File	Contents
front.mp4	N video images
states.json	N+1 states; actual/input global camera poses and actor states
actions.json	N transitions; vehicle controls and actual/input relative camera poses
collisions.json	Collision sensor impulses, simulator frames and contacted actors
chunks.npz	Original eight-action (0.5 s) chunks, relative poses and padding masks
metadata.json	Camera calibration, simulation settings, seed and pose definitions
validation.json	Collection validation results
coverage.json	Observed pre-contact road cells
_SUCCESS.json	Accepted-sample marker
contact_sheet.jpg	Source preview
pose_pair.png	Optional visualization, present in 83 preserved samples

Video image i corresponds to state i. Action i maps state i to state i+1. T_world_camera is actual c2w; T_world_input_camera is constructed input c2w. Camera axes are right/down/forward; world coordinates use the right-handed convention obtained by negating CARLA Y. Translations are in meters. Intrinsics are [[640, 0, 640], [0, 640, 352], [0, 0, 1]] at the native resolution. Per-sample metadata contains the exact conventions and calibration.

Input poses are constructed conditioning, not counterfactual collision-free ground truth. Before the first affected transition they equal actual poses. For first contact state j, the affected transition begins at j-1. Subsequent input poses extrapolate a fixed body-motion increment fitted to the preceding eight measured transitions, j-9 through j-2, anchored at state j-1. The fit does not use post-impact motion. Collection requires a stable pre-contact fit and at least 0.25 m displacement difference over the checked 0.5-second horizon.

Collision impulses identify events, not continuous contact intervals. physical_contact_end_s and physical_contact_duration_s are null. Post-impact observation duration is not contact duration.

Five-Second Windows

For a source with N images, window k has end = N - 1 - 80*k and start = end - 80. Keep only start >= 0; k=0 is the tail window. Bounds are inclusive: read images and states with [start:end+1] and actions with [start:end]. Each window contains 81 images and 80 transitions, spanning exactly 5 seconds between endpoint images. Its playback duration at 16 FPS is 5.0625 seconds. Adjacent windows share one endpoint image, but no action interval. The incomplete prefix is discarded. Final source state N and action N-1 do not have an endpoint image and are excluded from the window index.

Of the 2,248 sources, 2,155 contribute windows. The 93 excluded sources remain available in the archives: 89 have fewer than 80 images and four have exactly 80. Each included source contributes exactly one window containing first collision and zero, one or two earlier windows. contains_first_collision tests start < first_contact_state <= end. The entire tail window is not necessarily in contact. Event indices refer to the source collisions.json; first_collision_time_in_clip_s is relative to window start.

Captions

All 3,764 scene_static captions are complete. The model received every image in each window as native-resolution lossless PNG, in order, with server frame resampling disabled. The task was to describe the visible static environment; action, collision chronology and camera motion were excluded from the prompt. Source IDs, behavior labels and pose annotations were not supplied to the model.

Captions are model-generated text with JSON/consistency checks, not human-verified semantic ground truth. The release preserves the original wording, including any model mistakes. See the annotation method for exact settings, provenance and reproduction commands.

Distribution and Limitations

Behavior family	Source videos
catch_slow_vehicle	316
crossing_from_left	223
crossing_from_right	218
cut_in_from_left	194
cut_in_from_right	202
turn_then_wall	190
wall_on_left	474
wall_on_right	431

The ten maps are Town01 (15), Town02 (1), Town03 (11), Town05 (15), Town06 (5), Town10HD (2), Town11 (32), Town12 (833), Town13 (1,281), and Town15 (53). Town12 and Town13 account for approximately 94% of source videos. Counts include the 93 sources excluded from the window tables. Per-window map counts and exclusion reasons are available in clips_5s.json.

The September 9, 2026 collection comprises 83 preserved and 2,165 newly accepted samples. Each source is at most 20 seconds and has at most four seconds after first contact. There is no recovery phase or independent set of safe-driving controls; pre-collision windows originate from eventual collision encounters. Interactions are deliberately constructed, not naturalistic accident frequencies. Town11, Town12 and Town13 omit pedestrians. Reverse-driving and U-turn families are absent. Collection validation does not establish real-world transfer or independent physical-contact label accuracy. Full counts and source code hashes are preserved in the original manifests and per-sample metadata.

License and Attribution

Dataset files, captions and documentation are released under CC BY 4.0; the full legal text is in LICENSE. Code under tools/ is Apache-2.0, with its license in tools/LICENSE. These grants cover the contributors' rights in this release; upstream software and assets retain their existing notices.

Attribute this dataset by its title, the repository's listed creator(s), and the URL and revision of the repository you obtained it from. Include the CC BY 4.0 license link and indicate modifications. See NOTICE for component provenance. No repository ID is assumed by this local release.