SOIM: a self-organizing invertible map with applications in active vision

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Active vision systems require robust spatial representations for target localization.
• Learning invariant representations is challenging due to dynamic camera configurations.

Purpose of the Study:

• To propose a novel neural network, the Self-Organized Invertible Map (SOIM), for learning many-to-one functional mappings.
• To develop an invariant spatial representation for 3D point targets in active vision.
• To demonstrate the utility of SOIM in real-world active vision applications.

Main Methods:

• Development of the Self-Organized Invertible Map (SOIM) neural network architecture.
• Online learning of many-to-one functional mappings for spatial representation.
• Derivation and experimental verification of convergence and invariance properties.
• Implementation on a real active vision system.

Main Results:

• The SOIM successfully learned a many-to-one functional mapping for spatial representation of 3D point targets.
• The learned spatial representation demonstrated invariance to changing camera configurations.
• The invertible property of SOIM was validated and shown to be beneficial for active vision tasks.
• Convergence during learning was mathematically proven and experimentally confirmed.

Conclusions:

• The SOIM offers an efficient and feasible method for learning invariant spatial representations in active vision.
• The SOIM's properties are advantageous for various active vision applications, enhancing robotic perception and control.
• This work contributes a novel neural network approach to address key challenges in dynamic spatial understanding for robotic systems.