Continuous Viewpoint Planning in Conjunction with Dynamic Exploration for Active Object Recognition.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Active object recognition (AOR) improves performance by adjusting an agent's viewpoint.
• Viewpoint planning (VP) is crucial for AOR, but discrete methods suffer from quantization errors.
• Existing VP methods often involve sampling the viewpoint space, leading to inefficiencies.

Purpose of the Study:

• To propose a novel continuous viewpoint planning (VP) approach for active object recognition (AOR).
• To address the limitations of discrete viewpoint exploration in AOR.
• To enhance recognition performance through optimized viewpoint selection.

Main Methods:

• Utilized reinforcement learning (RL) for continuous VP.
• Employed two neural networks to model the VP policy as a parameterized Gaussian distribution.
• Applied proximal policy optimization (PPO) for policy learning.
• Introduced an adaptive entropy regularization for dynamic exploration.

Main Results:

• The proposed continuous VP method demonstrated superior performance compared to existing approaches.
• Experimental results on the GERMS dataset validated the effectiveness of the approach.
• The adaptive exploration scheme effectively adjusted viewpoint exploration ability during learning.

Conclusions:

• The continuous VP approach offers a significant advancement for active object recognition.
• Reinforcement learning provides a robust framework for optimizing viewpoint planning.
• The proposed method effectively mitigates quantization errors inherent in discrete VP techniques.