Dataset Card for KITScenes-LongTail Videos

A FiftyOne grouped video dataset built from KIT-MRT/KITScenes-LongTail, a long-tail autonomous-driving benchmark for VLMs/VLAs.

Each scenario is a ~9 s, six-camera surround-view event with a high-level driving instruction, multiple candidate future trajectories, and multilingual expert reasoning traces.

This is a FiftyOne dataset with 103 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub
from huggingface_hub import snapshot_download


# Download the dataset snapshot to the current working directory

snapshot_download(
    repo_id="Voxel51/KITScenes-LongTail", 
    local_dir=".", 
    repo_type="dataset"
    )

# Load dataset from current directory using FiftyOne's native format
dataset = fo.Dataset.from_dir(
    dataset_dir=".",  # Current directory contains the dataset files
    dataset_type=fo.types.FiftyOneDataset,  # Specify FiftyOne dataset format
    name="KITScenes-LongTail"  # Assign a name to the dataset for identification
)

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

KITScenes-LongTail targets generalization to rare ("long-tail") driving events for end-to-end driving. It provides synchronized multi-view video, ego trajectories, high-level instructions, and detailed multilingual reasoning traces, supporting in-context learning and few-shot generalization for multimodal models (VLMs and VLAs). Beyond safety/comfort metrics, it is designed to evaluate instruction following and the semantic coherence between a model's reasoning and its predicted trajectory. The reasoning traces are authored by domain experts (self-driving researchers) with diverse cultural backgrounds in English, Spanish, and Chinese.

This FiftyOne version models each scenario as one group with six camera video slices, a composite surround slice, and a 3D point-cloud slice reconstructed from the front camera. The observed clip (~4 s) is real footage; where labels exist (train split), the 5 s prediction horizon is represented and the expert reasoning's action windows are stored as temporal detections (see Dataset Structure for the exact, faithful breakdown of what is verbatim vs. derived vs. synthetic). It also ships with ready-made saved views and Qwen3-VL clip embeddings (with similarity, 2D visualization, uniqueness, and representativeness indexes) for exploration.

• Curated by: Institute of Measurement and Control Systems (MRT), Karlsruhe Institute of Technology (KIT), with collaborators from FZI Research Center for Information Technology, University Charles III of Madrid, Technical University of Madrid, University of Toronto, and Delft University of Technology.
• Funded by: German Federal Ministry for Economic Affairs and Energy, project "NXT GEN AI METHODS"; compute provided by NHR@KIT (HoreKa).
• Shared by: Original dataset by KIT-MRT; FiftyOne conversion published at harpreetsahota/KITScenes-LongTail.
• Language(s): English, Spanish, Chinese (reasoning traces).
• License: CC BY-NC 4.0, with additional dataset terms (non-commercial; cite the publication; anonymized faces/plates; data provided "as is"). Where the dataset terms conflict with CC BY-NC 4.0, the dataset terms prevail.

Dataset Sources

• Repository: https://huggingface.co/datasets/KIT-MRT/KITScenes-LongTail (original)
• Paper: Wagner et al., "LongTail Driving Scenarios with Reasoning Traces: The KITScenes LongTail Dataset," arXiv:2603.23607 (2026)

Uses

Direct Use

• Benchmarking VLMs/VLAs for end-to-end driving in long-tail scenarios.
• Trajectory prediction: forecasting a 5 s future trajectory (25 waypoints @ 5 Hz) from the past 4 s, the surround video, and a high-level instruction.
• Instruction-following evaluation with fine-grained commands (e.g. "overtake truck driving on the right").
• Studying semantic coherence between reasoning traces and predicted trajectories.
• Zero-shot, few-shot, and few-shot chain-of-thought (CoT) prompting; multilingual reasoning research across English/Spanish/Chinese.
• Multi-camera (360°) video understanding and qualitative review via the synchronized surround montage slice.
• Dataset exploration with the bundled Qwen3-VL clip embeddings: similarity / nearest-neighbour search, 2D embedding plots, and uniqueness / representativeness ranking to surface the rarest or most prototypical scenarios.
• Inspecting per-frame monocular depth maps and the merged front-camera 3D point cloud (the threed slice) in the FiftyOne 3D viewer.

Out-of-Scope Use

• Commercial use (the license is non-commercial).
• Perception tasks needing ground-truth object annotations (2D/3D boxes, masks, tracks, depth, HD maps): none are provided. The threed point cloud and per-frame depth maps are model-estimated (VGGT-Omega), not sensor-measured ground truth, so they should not be used as depth/geometry labels.
• Photometric or geometric tasks on the synthetic future-horizon frames of train videos (these freeze the last observed frame and are not real imagery).
• Treating the held-out test split as labeled: its future trajectories and reasoning are withheld.

Dataset Structure

Topology. media_type = group. 103 groups (3 train + 100 test), each with 8 group slices (seven video, one 3D):

• front_left, front, front_right, rear_left, rear, rear_right — the six surround cameras (one H.264 mp4 each).
• surround — a derived 2×3 montage of the six cameras playing in sync (top row: front-left/front/front-right; bottom row: rear-left/rear/rear-right).
• threed — a 3D point-cloud scene (.fo3d, media_type = 3d) reconstructed from the front clip, viewable in the FiftyOne 3D viewer (see 3D reconstruction below).

Each sample is also tagged with its split (train / test), and a split string field is stored for convenience.

Timeline. Videos are 5 Hz. Frames 1..n_observed (n_observed = 21, ~4 s) are the real observed clip. On train, frames after n_observed freeze the last observed frame to represent the 5 s prediction horizon, so the reasoning action windows can be stored as TemporalDetections with valid frame supports (train clips are 46 frames; test clips are 21 frames, observation-only).

Sample fields

Field	FiftyOne type	Description
scenario_id	StringField	Scenario identifier, e.g. "0000400" (verbatim)
split	StringField	"train" or "test" (also a sample tag)
camera	StringField	Slice name for this sample (a camera, or surround)
fps	IntField	Clip frame rate (5 Hz; from the paper)
n_observed_frames	IntField	Number of real observed frames (21)
n_future_frames	IntField	Synthetic future-horizon frames (25 on train, 0 on test)
now_frame	IntField	1-indexed frame marking "now" (the last observed frame)
driving_instruction	Classification	High-level maneuver instruction (verbatim)
scenario_type	Classification	Long-tail scenario category (verbatim)
trajectory_bev	DictField	Ego BEV waypoints [x, y] (x forward, y left, meters): past (21 pts) plus populated future variants (verbatim values; [[-100,-100]] placeholders dropped)
reasoning_english / reasoning_spanish / reasoning_chinese	DictField	The 9 expert reasoning fields per language, verbatim (train only)
driving_actions_english / _spanish / _chinese	TemporalDetections	Acceleration/steering action windows (0–3 s, 3–5 s); labels/reasons verbatim, frame supports derived (train only)
qa_english_reconstructed	ListField	Reconstructed 5 Q&A pairs (paper's English question templates + verbatim answers) (train only)
reasoning_cot_english_reconstructed	StringField	Reconstructed chain-of-thought block (train only)
scene_3d	StringField	Path to the front-camera .fo3d 3D scene (also materialized as the threed slice)
qwen_emb_surround / qwen_emb_front / qwen_emb_rear	VectorField	2048-d Qwen3-VL clip embedding (one per slice; on the surround / front / rear slices respectively)
uniqueness_surround / uniqueness_front / uniqueness_rear	FloatField	Per-clip uniqueness score in [0, 1] (higher = more unusual within its slice)
representativeness_surround / representativeness_front / representativeness_rear	FloatField	Per-clip representativeness score in [0, 1] (higher = more prototypical within its slice)

Frame fields (attached to the six camera slices only; not the surround montage)

Field	FiftyOne type	Description
past_trail	Polylines	History breadcrumb; grows over observed frames (per-frame ego-pose transform), then persists through the horizon
traj_expert_like	Polylines	Expert (GT) future plan, progressively built out across the horizon
traj_wrong_speed	Polylines	Adversarial future (wrong speed), built out across the horizon
traj_off_road	Polylines	Adversarial future (off road), built out across the horizon
traj_crash	Polylines	Adversarial future (crash), where present
plan_cursor	Keypoints	Single marker advancing along the expert plan

Trajectories are rendered as Polylines (connected paths) rather than Keypoints because a driving trajectory is an ordered, connected curve; the advancing plan_cursor remains a single Keypoint.

The front slice additionally carries per-frame outputs from the 3D reconstruction (sampled at 5 Hz):

Field	FiftyOne type	Description
depth_map	Heatmap	Monocular depth map for the frame, rendered as an overlay in the App
camera_translation	ListField	Camera position (3-vector) in the scene's OpenCV world frame
camera_rotation_matrix	ListField	Camera rotation as a 3×3 matrix
camera_quaternion	ListField	Camera rotation as a quaternion [qw, qx, qy, qz]
intrinsic_matrix	ListField	Recovered 3×3 pinhole intrinsics for the frame
fov_h_deg / fov_w_deg	FloatField	Estimated horizontal / vertical field of view, in degrees

dataset.info

• camera_parameters — per-camera pinhole K and extrinsics R, t (ego→camera optical frame), copied verbatim from the dataset card; the rig is fixed so they apply to every sample.
• camera_parameters_note — explanation of the calibration convention.
• timeline_note — describes the 5 Hz timeline and the synthetic freeze-frame horizon.
• provenance — explicit classification of every field as verbatim_from_dataset, derived, or constructed_or_synthetic, plus assumptions (ground_z = -1.6 m, fps = 5 Hz, synthetic horizon).

Parsing decisions

• Verbatim: scenario_id, driving_instruction, scenario_type, the per-language reasoning_* dicts, trajectory_bev waypoint values, and the real observed frames.
• Derived: trajectory overlays are projected with the card's camera matrices plus a road-plane estimate ground_z = -1.6 m (and estimated per-frame headings for the past trail); fps/timing are taken from the paper (the parquet has no timestamps); temporal-detection frame supports map the paper's 0–3 s/3–5 s windows onto the horizon.
• Constructed/synthetic: the train future-horizon frames (a freeze of the last real frame); qa_english_reconstructed / reasoning_cot_english_reconstructed (assembled from paper question templates + verbatim answers). Spanish/Chinese Q&A are intentionally not fabricated.
• Held-out: on test, future trajectories and all reasoning are withheld (empty), so those samples are inputs-only (video + instruction + scenario_type + past trajectory), have no future horizon, and no temporal detections.
• Scenario-level labels are attached to every slice (including surround); per-pixel overlays are attached only to the camera slices (their coordinates are per-view, not montage-space).

Saved views

The dataset ships with these saved views (available from the view dropdown in the App or via dataset.load_saved_view(...)):

Saved view	Description
slice__front_left, slice__front, slice__front_right, slice__rear_left, slice__rear, slice__rear_right	One per camera — flattens the group to a single camera so you can scan all 103 clips from that viewpoint
by_scenario_type_and_instruction	The surround clips dynamically grouped by scenario_type, ordered by driving_instruction within each group

Embeddings, similarity & quality scores

Each of the surround, front, and rear slices carries a 2048-d Qwen3-VL clip embedding (qwen_emb_<slice>) summarizing the whole video clip, along with FiftyOne Brain indexes computed from those embeddings:

Brain key	Type	Use
sim_surround, sim_front, sim_rear	similarity	Nearest-neighbour / sort-by-similarity search over clips
viz_surround, viz_front, viz_rear	visualization (UMAP)	2D embedding scatter plot for interactive exploration
uniqueness_surround, uniqueness_front, uniqueness_rear	uniqueness	Per-clip uniqueness_<slice> score (find rare / outlier clips)
representativeness_surround, representativeness_front, representativeness_rear	representativeness	Per-clip representativeness_<slice> score (find prototypical clips)

3D reconstruction

The threed slice contains one .fo3d scene per scenario, reconstructed from the front clip with VGGT-Omega. Each scene holds a merged, colored point cloud plus camera frustums for the sampled frames and opens in the FiftyOne 3D viewer. The matching per-frame depth maps and camera poses live on the front slice's frame fields (see the frame-fields table above). The point cloud and depth are model-estimated (not sensor LiDAR), and are derived from the real observed frames.

Dataset Creation

Curation Rationale

Generalization in perception has advanced, but decision-making in long-tail scenarios remains a major challenge. The dataset couples driving with high-level instructions and multilingual, step-by-step reasoning traces to accelerate progress on decision-making in rare events, and to enable studying how reasoning style and language affect driving competence. Multiple plausible maneuvers are evaluated rather than a single expert trajectory (the paper introduces a multi-maneuver score, MMS).

Source Data

Data Collection and Processing

Recorded over two years beginning in late 2023 across urban, suburban, and highway environments (main locations: Karlsruhe, Heidelberg, Mannheim, and the Black Forest). Routes were adjusted to include construction zones and intersections, and filtered for rare events such as adverse weather (heavy rain, snow, fog), road closures, and accidents. Long-tail data was further selected using the Pareto principle with nuScenes as a reference distribution (80% cumulative-frequency threshold on rank-frequency plots). The full collection is 1,000 nine-second scenarios split train (500) / test (400) / validation (100); the released subset used here includes the test split and a small set of train samples. Each scenario provides synchronized six-view video over a 360° horizontal field of view at 5 Hz, ~4 s observed, in raw, pinhole, and stitched formats (this FiftyOne build uses the pinhole frames, downscaled for encoding). The future prediction horizon is 5 s (25 waypoints at 5 Hz).

Who are the source data producers?

The KIT-MRT team and collaborators, recording with a research vehicle sensor suite.

Annotations

Annotation process

High-level driving instructions were manually annotated by domain experts. Reasoning traces were collected by asking experts five questions per scenario: one open-ended situational-awareness question grounded in the instruction, and four questions about the acceleration and steering decisions for the next 0–3 s and the final 3–5 s of the expert trajectory. Action labels follow heuristics (acceleration: slight/strong accelerate, decelerate, maintain; steering: slight/sharp left/right, straight). Experts answered in their mother tongue (or a fluent language); verbal responses were transcribed with Whisper. For evaluation, reference trajectories are provided per scenario across categories (expert-like, wrong speed, neglect instruction, off road, crash); expert trajectories are driven, wrong-speed variants are augmented via state estimation and spline modification, and the remaining categories are manually labeled.

Who are the annotators?

Domain experts (researchers working on self-driving) with diverse linguistic and cultural backgrounds (English, Spanish, Chinese).

Personal and Sensitive Information

Faces and license plates are anonymized using state-of-the-art anonymization software (BrighterAI). Requests for removal of specific frames can be directed to info@mrt.kit.edu.

Citation

BibTeX:

@misc{wagner2026longtaildrivingscenariosreasoning,
  title={LongTail Driving Scenarios with Reasoning Traces: The KITScenes LongTail Dataset},
  author={Royden Wagner and Omer Sahin Tas and Jaime Villa and Felix Hauser and Yinzhe Shen and
          Marlon Steiner and Dominik Strutz and Carlos Fernandez and Christian Kinzig and
          Guillermo S. Guitierrez-Cabello and Hendrik Königshof and Fabian Immel and Richard Schwarzkopf and
          Nils Alexander Rack and Kevin Rösch and Kaiwen Wang and Jan-Hendrik Pauls and Martin Lauer and
          Igor Gilitschenski and Holger Caesar and Christoph Stiller},
  year={2026},
  eprint={2603.23607},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.23607},
}

APA:

Wagner, R., Taş, Ö. Ş., Villa, J., Hauser, F., Shen, Y., Steiner, M., Strutz, D., Fernandez, C., Kinzig, C., Guitierrez-Cabello, G. S., Königshof, H., Immel, F., Schwarzkopf, R., Rack, N. A., Rösch, K., Wang, K., Pauls, J.-H., Lauer, M., Gilitschenski, I., Caesar, H., & Stiller, C. (2026). LongTail Driving Scenarios with Reasoning Traces: The KITScenes LongTail Dataset. arXiv:2603.23607.

More Information

The released subset contains the test split and a small number of train samples (few-shot examples); the full train and validation splits, along with stitched 360° images, are announced for a later dataset version.

Dataset Card Contact

For the underlying dataset: info@mrt.kit.edu