README.md

metadata

language:
  - en
license: cc-by-nc-sa-4.0
task_categories:
  - visual-question-answering
  - object-detection
  - image-text-to-text
pretty_name: GRAID BDD100K Question-Answer Dataset
tags:
  - visual-reasoning
  - spatial-reasoning
  - object-detection
  - computer-vision
  - autonomous-driving
  - bdd100k
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
dataset_info:
  features:
    - name: image
      dtype: image
    - name: annotations
      list:
        - name: area
          dtype: float64
        - name: bbox
          list: float64
        - name: category
          dtype: string
        - name: category_id
          dtype: int64
        - name: iscrowd
          dtype: int64
        - name: score
          dtype: float64
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: question_type
      dtype: string
    - name: source_id
      dtype: string
    - name: id
      dtype: int64
  splits:
    - name: train
      num_bytes: 1153019078219.5
      num_examples: 4627468
    - name: val
      num_bytes: 167621242099.75
      num_examples: 672330
  download_size: 30647805316
  dataset_size: 1320640320319.25

GRAID BDD100K Question-Answer Dataset

Paper | Project Page

Overview

This dataset was generated using GRAID (Generating Reasoning questions from Analysis of Images via Discriminative artificial intelligence), a framework for creating spatial reasoning datasets from object detection annotations.

GRAID transforms raw object detection data into structured question-answer pairs that test various aspects of object localization, visual reasoning, spatial reasoning, and object relationship comprehension.

Dataset Details

• Total QA Pairs: 5,299,798
• Source Dataset: BDD100K (Berkeley DeepDrive)
• Generation Date: 2025-09-11
• Image Format: Embedded in parquet files (no separate image files)
• Question Types: 22 different reasoning patterns

Dataset Splits

• train: 4,627,468 (87.31%)
• val: 672,330 (12.69%)

Question Type Distribution

• Is there at least one {object_1} to the left of any {object_2}?: 708,762 (13.37%)
• Is there at least one {object_1} to the right of any {object_2}?: 708,762 (13.37%)
• Is there at least one {object_1} that appears closer to the camera than any {object_2}?: 708,762 (13.37%)
• Is there at least one {object_1} that appears farther from the camera than any {object_2}?: 708,762 (13.37%)
• Are there more than {target} {object_1}(s) in this image? Respond Yes/No.: 510,268 (9.63%)
• Are there less than {target} {object_1}(s) in this image? Respond Yes/No.: 510,268 (9.63%)
• Are there more {object_1}(s) than {object_2}(s) in this image?: 305,574 (5.77%)
• How many {object_1}(s) are there in this image?: 255,134 (4.81%)
• What appears the most in this image: {object_1}s, {object_2}s, or {object_3}s?: 250,197 (4.72%)
• How many {object_1}(s) are in the image? Choose one: A) {range_a}, B) {range_b}, C) {range_c}, D) Unsure / Not Visible. Respond with the letter only.: 136,928 (2.58%)
• Rank the {k} kinds of objects that appear the largest (by pixel area) in the image from largest to smallest. Provide your answer as a comma-separated list of object names only.: 111,176 (2.10%)
• Divide the image into a grid of {N} rows x {M} columns. Number the cells from left to right, then top to bottom, starting with 1. In what cell does the {object_1} appear?: 82,442 (1.56%)
• What kind of object appears the most frequently in the image?: 64,405 (1.22%)
• Rank the {k} kinds of objects that appear the closest to the camera in the image from closest to farthest. Provide your answer as a comma-separated list of object names only.: 60,120 (1.13%)
• If you were to draw a tight box around each object in the image, which type of object would have the biggest box?: 57,994 (1.09%)
• What kind of object appears the least frequently in the image?: 53,454 (1.01%)
• Divide the image into thirds. In which third does the {object_1} primarily appear? Respond with the letter only: A) left third, B) middle third, C) right third.: 37,730 (0.71%)
• What is the rightmost object in the image?: 8,287 (0.16%)
• What is the leftmost object in the image?: 7,327 (0.14%)
• Is the width of the {object_1} appear to be larger than the height?: 7,059 (0.13%)
• Does the leftmost object in the image appear to be wider than it is tall?: 3,682 (0.07%)
• Does the rightmost object in the image appear to be wider than it is tall?: 2,705 (0.05%)

Performance Analysis

Question Processing Efficiency

Question Type	is_applicable Avg (ms)	apply Avg (ms)	Predicate -> QA Hit Rate	Empty cases
Divide the image into thirds. In which third does the {object_1} primarily appear? Respond with the letter only: A) left third, B) middle third, C) right third.	0.03	0.92	71.7%	11535
Is the width of the {object_1} appear to be larger than the height?	0.01	1.15	16.7%	34017
Divide the image into a grid of {N} rows x {M} columns. Number the cells from left to right, then top to bottom, starting with 1. In what cell does the {object_1} appear?	0.01	5.19	42.5%	93933
If you were to draw a tight box around each object in the image, which type of object would have the biggest box?	0.02	28.15	78.8%	15593
Rank the {k} kinds of objects that appear the largest (by pixel area) in the image from largest to smallest. Provide your answer as a comma-separated list of object names only.	0.02	26.79	87.0%	16663
What kind of object appears the most frequently in the image?	0.02	0.01	87.5%	9182
What kind of object appears the least frequently in the image?	0.01	0.01	72.6%	20133
Is there at least one {object_1} to the left of any {object_2}?	6.47	69.93	100.0%	0
Is there at least one {object_1} to the right of any {object_2}?	5.17	46.95	100.0%	0
What is the leftmost object in the image?	0.03	2.19	18.0%	33486
What is the rightmost object in the image?	0.02	2.12	20.3%	32526
How many {object_1}(s) are there in this image?	0.02	0.02	100.0%	0
Are there more {object_1}(s) than {object_2}(s) in this image?	0.01	0.02	97.7%	1708
What appears the most in this image: {object_1}s, {object_2}s, or {object_3}s?	0.01	0.02	69.5%	22432
Does the leftmost object in the image appear to be wider than it is tall?	0.01	1.62	9.0%	37131
Does the rightmost object in the image appear to be wider than it is tall?	0.01	1.43	6.6%	38108
Are there more than {target} {object_1}(s) in this image? Respond Yes/No.	0.01	0.02	100.0%	0
Are there less than {target} {object_1}(s) in this image? Respond Yes/No.	0.01	0.02	100.0%	0
How many {object_1}(s) are in the image? Choose one: A) {range_a}, B) {range_b}, C) {range_c}, D) Unsure / Not Visible. Respond with the letter only.	0.01	0.14	94.3%	4504
Is there at least one {object_1} that appears closer to the camera than any {object_2}?	4.09	8627.26	100.0%	0
Is there at least one {object_1} that appears farther from the camera than any {object_2}?	1.92	420.61	100.0%	0
Rank the {k} kinds of objects that appear the closest to the camera in the image from closest to farthest. Provide your answer as a comma-separated list of object names only.	0.04	54.12	47.0%	67719
Notes:

• is_applicable checks if a question type can be applied to an image
• apply generates the actual question-answer pairs
• Predicate -> QA Hit Rate = Percentage of applicable cases that generated at least one QA pair
• Empty cases = Number of times is_applicable=True but apply returned no QA pairs

Usage

from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("kd7/graid-bdd")

# Access individual splits
train_data = dataset["train"]
val_data = dataset["val"]

# Example of accessing a sample
sample = dataset["train"][0]  # or "val"
print(f"Question: {sample['question']}")
print(f"Answer: {sample['answer']}")  
print(f"Question Type: {sample['question_type']}")

# The image is embedded as a PIL Image object
image = sample["image"]
image.show()  # Display the image

Dataset Schema

• image: PIL Image object (embedded, no separate files)
• annotations: COCO-style bounding box annotations
• question: Generated question text
• answer: Corresponding answer text
• reasoning: Additional reasoning information (if applicable)
• question_type: Type of question (e.g., "HowMany", "LeftOf", "Quadrants")
• source_id: Original image identifier from BDD100K (Berkeley DeepDrive)

License

This dataset is derived from the BDD100K dataset. Please refer to the BDD100K license terms for usage restrictions. The GRAID-generated questions and metadata are provided under the same terms.

Citation

If you use this dataset in your research, please cite both the original dataset, the BDD100K dataset, and the GRAID framework paper:

@article{graid2025spatial,
  title={GRAID: Enhancing Spatial Reasoning of VLMs Through High-Fidelity Data Generation},
  author={{Anonymous}},
  journal={arXiv preprint arXiv:2510.22118},
  year={2025},
  url={https://huggingface.co/papers/2510.22118}
}

@dataset{graid_bdd,
    title={GRAID BDD100K Question-Answer Dataset},
    author={GRAID Framework},
    year={2025},
    note={Generated using GRAID: Generating Reasoning questions from Analysis of Images via Discriminative artificial intelligence}
}

@INPROCEEDINGS{9156329,
    author={Yu, Fisher and Chen, Haofeng and Wang, Xin and Xian, Wenqi and Chen, Yingying and Liu, Fangchen and Madhavan, Vashisht and Darrell, Trevor},
    booktitle={2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, 
    title={BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning}, 
    year={2020},
    volume={},
    number={},
    pages={2633-2642},
    keywords={Task analysis;Visualization;Roads;Image segmentation;Meteorology;Training;Benchmark testing},
    doi={10.1109/CVPR42600.2020.00271}
}

Contact

For questions about this dataset or the GRAID framework, please open an issue in the repository.