2D shape dataset of proximity-based spatial relation concepts

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual perception relies on understanding spatial relationships between objects.
• Teaching AI these spatial concepts is challenging, especially with limited data.
• Few-shot and meta-learning methods are crucial for AI to learn from sparse examples.

Purpose of the Study:

• To create a dataset supporting AI research in learning basic 2D spatial concepts.
• To facilitate few-shot and meta-learning approaches for spatial relation recognition.
• To provide a resource for AI models to generalize to new spatial concepts with minimal data.

Main Methods:

• Dataset generation using a Python script for creating 2D images with one or two objects.
• Inclusion of diverse object types: regular geometric shapes, real-world segmented objects, and structural anomalies.
• Varied object attributes: shape, size, positioning, and spatial relationships (alone, close, far, overlap).

Main Results:

• A lightweight, modular dataset enabling rapid experimentation in AI spatial concept learning.
• Pre-training capability for AI models on simple geometric concepts.
• Enhanced model generalization to novel classes using minimal labeled examples.

Conclusions:

• The dataset effectively supports AI in learning fundamental spatial concepts, particularly in low-data regimes.
• It is applicable to domains requiring geometric relationship analysis with limited annotated samples, such as electronic component inspection.
• The dataset's design promotes efficient AI model development and improved performance in recognizing spatial patterns.