Dataset¶
The 24 conformance scenarios are published on Hugging Face as mehmetkeremturkcan/bikeped, with per-frame ground-truth labels and three videos per scenario. This page describes what the dataset contains and how to read it.
Contents¶
The dataset card names it "bikeped Conformance Scenarios". Agents follow hand-authored waypoint paths that are interpolated at 30 fps, and each frame carries a danger label from the kinematic safety model. The licence is CC BY 4.0.
| Path | Files | Contents |
|---|---|---|
scenarios.jsonl |
1 | One row per scenario: metadata, waypoint paths and video paths. |
frames.jsonl |
1 | One row per scenario, agent and frame, with position and labels. |
videos/sim/ |
24 | Schematic simulated render of each scenario. |
videos/sim_2cam/ |
24 | A second schematic set. The folder name matches the _2cam suffix that sim_visualizer.py --two-cam gives its output. |
videos/carla/ |
24 | CARLA 0.9.16 renders from the pole camera on Town10HD_Opt. |
Videos are named by index and slug, for example videos/carla/07_fast_approach.mp4. The dataset page reports a total file size of 389 MB.
Configurations¶
The two JSONL files are exposed as two configurations with a single train split each.
scenarios¶
24 rows.
| Column | Type | Meaning |
|---|---|---|
scenario_id |
int | Index of the scenario, 0 to 23, in the order of decision_testbench.SCENARIOS. |
name |
string | Display name, for example Fast Approach. |
expected_dominant |
string | idle, safe, warning or alert. |
vehicle |
string | Braking profile of the cyclist: bicycle or ebike. |
duration_s |
float | Scenario length in seconds. |
n_person, n_bike |
int | Number of pedestrians and cyclists. |
agents |
list | One entry per agent with type, agent_id and path, a list of {x, y, t} waypoints. |
sim_video, carla_video |
string | Paths of the schematic and the CARLA video inside the dataset. |
frames¶
12,901 rows.
| Column | Type | Meaning |
|---|---|---|
scenario_id, scenario_name |
int, string | Scenario the row belongs to. |
frame_idx, time_s |
int, float | Frame index and time at 30 fps. |
agent_id, agent_type |
int, string | Agent the position belongs to. |
x_m, y_m |
float | Interpolated position in the scenario frame, in metres. |
is_danger |
bool | Ground-truth danger label of the frame. |
actionable |
bool | Whether a warning at this frame can still be acted on. |
severity |
float | Severity, scaled by speed relative to 12 m/s. |
ttc_s |
float | Time to collision in seconds. |
closing_speed_ms, bike_speed_ms |
float | Closing speed of the pair and cyclist speed, in m/s. |
stopping_dist_m |
float | Cyclist stopping distance in metres. |
warning_budget_s |
float | Warning budget in seconds. |
distance_m |
float | Distance to the nearest counterpart in metres. |
Coordinate frame¶
Positions use the scenario frame of the testbench. The crosswalk runs along the y axis at x = 5 m and is about 3 m wide, cyclists travel along x, and the card places the deployed camera at (x = 5, y = -13, z = 3.66) in metres.
Ground-truth labels¶
The card describes the label as a conservative kinematic model: a frame is dangerous when the pair is closing, the closest approach is within 5 m, and either the stopping distance or the time to collision says the cyclist cannot avoid the conflict. Conventional bicycles use a reaction time of 0.84 s and a deceleration of 1.96 m/s², and e-bikes a deceleration of 6.0 m/s². 21 of the 24 scenarios contain a danger interval. The same model is implemented by compute_ground_truth_danger in decision_testbench.py, documented under safety model.
Loading¶
Relation to the repository¶
The waypoint paths in scenarios.jsonl are those of decision_testbench.SCENARIOS, so a result computed from the dataset can be checked against python decision_testbench.py. The repository also renders the videos: sim_visualizer.py writes the schematic ones and the CARLA renderer the photorealistic ones.