Cross-Domain Indoor Visual Place Recognition for Mobile Robot via Generalization Using Style Augmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Visual recognition systems often struggle with domain shifts, such as changes in camera models or environments.
• Improving the robustness of scene classification across diverse indoor settings is crucial for autonomous systems.

Purpose of the Study:

• To develop an algorithm for robust multi-domain visual recognition of indoor environments.
• To enhance scene classification performance using synthetic and real-world data from various domains.
• To improve the performance of models on unseen domains through style randomization and transfer learning.

Main Methods:

• Utilized a convolutional neural network (CNN) architecture.
• Implemented style randomization techniques to bridge domain gaps.
• Employed a transfer learning approach with style extension for multi-domain scene classification.
• Created and utilized a dataset encompassing diverse indoor scenarios, camera models, and conditions.

Main Results:

• The proposed method achieved an average accuracy of 92.08% in multi-domain scene classification.
• Multi-domain data and style enhancement significantly improved model performance.
• The approach demonstrated superior results compared to a previously reported method.

Conclusions:

• The developed algorithm effectively addresses the challenge of domain variation in indoor visual recognition.
• Style randomization and multi-domain data are key to enhancing the generalization capabilities of scene classification models.
• The findings have implications for improving the performance of humanoid robots in complex indoor environments.