Hanoi0126/visual-object-attributes

Dataset Description

This dataset contains two types of tasks:

Single-Object Tasks (5 tasks, 1,200 samples)

• Shape: Identify object shape (star, diamond, triangle, square)
• Color: Identify object color (red, green, blue, yellow)
• Size: Identify object size (xs, s, l, xl)
• Location: Identify object location (upper_left, upper_right, lower_left, lower_right)
• Angle: Identify triangle orientation (0°, 45°, 90°, 135°)

Multi-Object Tasks (3 tasks, 300 samples)

• Count: Count objects by color
• Position: Determine relative position (left, right, above, below)
• Occlusion: Identify which object is on top

Key Features

• White background for clarity
• Randomized attributes: Each task randomizes non-target attributes to prevent shortcuts
• Description field: Enables text-only experiments (image vs. text-only comparison)
• Balanced classes: Equal samples per class within each task

Dataset Structure

{
    "image": PIL.Image,
    "task": str,  # Task name
    "label": str,  # Ground truth label
    "description": str,  # Text description of the image
    "dataset_type": str,  # "single" or "multi"
    "question": str,  # Question text (multi-object only)
    "options": List[str],  # Answer options (multi-object only)
    "shape": str,  # Object shape (if applicable)
    "color": str,  # Object color (if applicable)
    "size_class": str,  # Size class (if applicable)
    "location": str,  # Location (if applicable)
    "angle": str,  # Angle (if applicable)
}

Usage

from datasets import load_dataset

# Load entire dataset
ds = load_dataset("Hanoi0126/visual-object-attributes")

# Load specific task
ds_color = load_dataset("Hanoi0126/visual-object-attributes", split="color")