nooranis/shape-counting-dataset

Shape Counting Dataset

A dataset for evaluating shape counting abilities in vision models and humans.

Dataset Description

This dataset contains images with varying numbers of squares, triangles, and stars on a white background. Each image is provided in multiple versions: the original clean image plus several noisy variants.

Image Specifications

• Size: 256×256 pixels
• Format: Grayscale PNG
• Shape size: 18 pixels
• Background: White (255)
• Shapes: Black (0)

Dataset Structure

Fields

Field	Type	Description
image	Image	The shape image (original or noisy)
image_id	int	Unique ID for the base image (same across noise variants)
noise_type	string	Type of noise applied (see below)
count_squares	int	Number of squares in the image
count_triangles	int	Number of triangles in the image
count_stars	int	Number of stars in the image
total_shapes	int	Total number of shapes (sum of all counts)
bucket	int	Bucket number (1, 2, or 3)
bucket_name	string	Bucket description
difficulty	string	"medium" or "hard"
intended_counts	string	Original intended counts before placement

Buckets Explained

The dataset is organized into 3 buckets based on shape type complexity:

Bucket	Name	Description
1	single_shape	Only ONE type of shape (all squares, all triangles, or all stars)
2	two_shapes	TWO different shape types mixed together
3	three_shapes	ALL THREE shape types mixed together

Difficulty Levels

Difficulty	Total Shapes	Description
medium	11-36	Moderate counting difficulty
hard	37-60	Challenging counting task

Noise Types

Each image is provided in 8 variants:

Noise Type	Description
original	Clean image, no noise
salt_pepper_medium	15% salt & pepper noise
salt_pepper_heavy	25% salt & pepper noise
salt_pepper_extreme	35% salt & pepper noise
blur_heavy	Gaussian blur (radius=3)
blur_extreme	Gaussian blur (radius=5)
motion_blur	Horizontal motion blur (size=9)
motion_blur_heavy	Horizontal motion blur (size=13)

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("nooranis/shape-counting-dataset")

# Filter for only original (clean) images
original_only = dataset.filter(lambda x: x['noise_type'] == 'original')

# Filter for a specific noise type
blurry = dataset.filter(lambda x: x['noise_type'] == 'blur_heavy')

# Filter by difficulty
hard_images = dataset.filter(lambda x: x['difficulty'] == 'hard')

# Filter by bucket
single_shape = dataset.filter(lambda x: x['bucket'] == 1)

Example: Get image and count

from datasets import load_dataset

dataset = load_dataset("nooranis/shape-counting-dataset")

# Get a sample
sample = dataset['train'][0]

# Access the image
image = sample['image']  # PIL Image

# Access counts
print(f"Squares: {sample['count_squares']}")
print(f"Triangles: {sample['count_triangles']}")
print(f"Stars: {sample['count_stars']}")
print(f"Total: {sample['total_shapes']}")

Example: Evaluate a vision model

from datasets import load_dataset

dataset = load_dataset("nooranis/shape-counting-dataset")

# Get original images only
test_data = dataset['train'].filter(lambda x: x['noise_type'] == 'original')

for sample in test_data:
    image = sample['image']
    true_count = sample['count_stars']  # or count_squares, count_triangles
    
    # Your model prediction here
    predicted = your_model.count_stars(image)
    
    # Compare
    is_correct = (predicted == true_count)

Dataset Statistics

• Total images: 1795 base images
• With noise variants: 14360 total rows
• Noise types: 8
• Buckets: 3
• Difficulty levels: 2 (medium, hard)