Unsupervised Active Visual Search With Monte Carlo Planning Under Uncertain Detections.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Active visual search requires efficient exploration strategies for object localization.
• Current methods often rely on supervised training and are sensitive to object detector failures.

Purpose of the Study:

• To develop an unsupervised active visual search solution robust to detector failures.
• To enhance exploration effectiveness using an intuitive probability distribution update mechanism.
• To improve the agent's belief update for more efficient searching.

Main Methods:

• Proposed POMP-BE-PD (Pomcp-based Online Motion Planning with Belief by Exploration and Probabilistic Detection) algorithm.
• Unsupervised learning approach requiring no training sessions.
• Integration of object detector success statistics into probability modeling.
• Utilizes POMDP solved via Monte-Carlo planning with agent pose and RGB-D observations.

Main Results:

• Achieved a 35% increase in average success rate across environments on the Active Vision Dataset Benchmark.
• Reduced average path length by 4% compared to competing methods.
• Demonstrated state-of-the-art performance without any prior training.

Conclusions:

• The POMP-BE-PD solution offers a more plausible and robust approach to active visual search.
• Unsupervised learning and probabilistic detection modeling significantly improve search efficiency and success rates.
• The method effectively handles object detector failures, enhancing real-world applicability.