Note

This is a community plugin, an external project maintained by its respective author. Community plugins are not part of FiftyOne core and may change independently. Please review each plugin’s documentation and license before use.

BDDOIA Safe/Unsafe Action Dataset for FiftyOne#

This dataset is designed for benchmarking safety-aware vision-language models within the FiftyOne ecosystem, particularly for autonomous driving safety evaluation. bddi-ezgif com-video-to-webp-converter

Dataset Overview#

Source & Citation#

Source: Based on the BDD-OIA dataset introduced in Xu et al., CVPR 2020
Original Paper: “Explainable Object-Induced Action Decision for Autonomous Vehicles”
License: Creative Commons Attribution 4.0 International

Dataset Statistics#

Split	Images	Annotations	Safe Actions	Unsafe Actions
Train	16,082	16,082	forward: 8,770, stop: 6,887, left: 218, right: 163, confuse: 16, unknown: 28	stop: 8,770, forward: 6,887, right: 218, left: 163, confuse: 16, unknown: 28
Validation	2,270	2,270	stop: 958, forward: 1,237, left: 39, right: 24, confuse: 6, unknown: 6	forward: 958, stop: 1,237, left: 24, right: 39, confuse: 6, unknown: 6
Test	4,572	4,572	forward: 2,484, stop: 1,965, left: 57, right: 51, confuse: 8, unknown: 7	stop: 2,484, forward: 1,965, right: 57, left: 51, confuse: 8, unknown: 7

Total: 22,924 images across all splits

Action Categories#

Each image has one safe action and one corresponding unsafe action:

Index	Action	Description
0	`forward`	Continue driving forward
1	`stop`	Stop the vehicle
2	`left`	Turn left
3	`right`	Turn right
4	`confuse`	Confused/uncertain situation

Safety Mapping#

The dataset creates unsafe scenarios by flipping actions:

forward ↔ stop (0↔1)
left ↔ right (2↔3)
confuse stays the same (4→4)

Explanation Classes (Reasons)#

Each annotation includes a multi-hot vector of 21 reasoning categories:

Index	Explanation Class	Train Count	Val Count	Test Count
0	Traffic light is green	5,474	752	1,574
1	Follow traffic	2,448	372	667
2	Road is clear	3,422	474	941
3	Traffic light	3,760	530	1,083
4	Traffic sign	1,095	141	302
5	Obstacle: car	170	22	39
6	Obstacle: person	116	13	34
7	Obstacle: rider	3,699	512	1,037
8	Obstacle: others	328	36	89
9	No lane on the left	100	21	29
10	Obstacles on the left lane	464	66	136
11	Solid line on the left	220	28	68
12	On the left-turn lane	115	16	23
13	Traffic light allows (left)	629	81	175
14	Front car turning left	270	26	69
15	No lane on the right	3,173	456	860
16	Obstacles on the right lane	3,171	438	897
17	Solid line on the right	2,582	359	713
18	On the right-turn lane	4,293	595	1,182
19	Traffic light allows (right)	2,791	396	826
20	Front car turning right	1,558	194	403

Quick Start#

Installation#

pip install fiftyone

Load Dataset via FiftyOne Zoo#

import fiftyone as fo
import fiftyone.zoo as foz

# Load the dataset (automatically downloads if not cached)
dataset = foz.load_zoo_dataset(
    "https://github.com/AdonaiVera/bddoia-fiftyone",
    split="validation",                                 # or "train", "test"
    max_samples=10000                                   # limit for quick exploration
)

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

Load Specific Split#

# Load training data
train_dataset = foz.load_zoo_dataset("https://github.com/AdonaiVera/bddoia-fiftyone", split="train")

# Load test data
test_dataset = foz.load_zoo_dataset("https://github.com/AdonaiVera/bddoia-fiftyone", split="test")

# Load all splits
all_splits = foz.load_zoo_dataset("https://github.com/AdonaiVera/bddoia-fiftyone")

Dataset Exploration#

View Dataset Statistics#

# Get dataset info
print(dataset)

# View sample annotations
sample = dataset.first()
print("Safe action:", sample.ground_truth.label)
print("Unsafe action:", sample.unsafe_action.label)
print("Reasons:", sample.reasons.labels)

Filter by Actions#

# View all forward actions
forward_samples = dataset.match(F("ground_truth.label") == "forward")
print(f"Found {len(forward_samples)} forward actions")

# View all stop actions
stop_samples = dataset.match(F("ground_truth.label") == "stop")
print(f"Found {len(stop_samples)} stop actions")

Filter by Reasons#

# View samples with traffic light issues
traffic_light_samples = dataset.match(
    F("reasons.labels").contains("Traffic light is green")
)
print(f"Found {len(traffic_light_samples)} samples with green traffic lights")

Analysis & Visualization#

Action Distribution#

# Plot action distributions
import fiftyone.core.plots as fop

# Safe actions
safe_plot = dataset.count("ground_truth.label")
safe_plot.show()

# Unsafe actions
unsafe_plot = dataset.count("unsafe_action.label")
unsafe_plot.show()

Reason Analysis#

# Most common reasons
reason_counts = dataset.count("reasons.labels")
print("Top reasons:", reason_counts[:10])

# Co-occurrence analysis
cooccurrence = dataset.count("reasons.labels", "ground_truth.label")
print("Action-Reason co-occurrence:", cooccurrence)

Custom Filters#

# Create custom view
custom_view = dataset.match(
    (F("ground_truth.label") == "forward") &
    (F("reasons.labels").contains("Traffic light is green"))
)

print(f"Found {len(custom_view)} samples")

References#

Original Paper: Xu et al., “Explainable Object-Induced Action Decision for Autonomous Vehicles”, CVPR 2020
Dataset: BDD-OIA Project
FiftyOne: Voxel51 Documentation
Remote Zoo: FiftyOne Remote Datasets

Contributing#

This dataset is designed for the workshop on “Benchmarking Safe Driving Perception with FiftyOne”. For questions or contributions:

License#

This dataset extension is released under the Creative Commons Attribution 4.0 International License, following the original BDD-OIA dataset license.

Note: This dataset is specifically designed for safety evaluation in autonomous driving scenarios. Always ensure proper validation and testing when using these annotations for real-world applications.