Structure-Aware Feature Disentanglement With Knowledge Transfer for Appearance-Changing Place Recognition.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Long-term visual place recognition (VPR) faces challenges due to drastic environmental appearance changes (time of day, season).
• Existing methods often neglect stable structural information, focusing on feature disentangling or style transfer.
• This limitation hinders robust performance in dynamic environments.

Purpose of the Study:

• To present a novel structure-aware feature disentanglement network (SFDNet) for enhanced long-term visual place recognition.
• To leverage stable structural information often overlooked by current VPR techniques.
• To improve the robustness and accuracy of VPR systems under extreme environmental variations.

Main Methods:

• Developed a structure-aware feature disentanglement network (SFDNet) incorporating knowledge transfer and adversarial learning.
• Employed probabilistic knowledge transfer (PKT) to integrate Canny edge detector knowledge into the structure encoder.
• Introduced an appearance teacher module to enrich appearance encoder learning beyond metric learning.

Main Results:

• The proposed SFDNet effectively utilizes structural information for image similarity measurement.
• Evaluated on six datasets with extreme environmental changes, demonstrating superior performance.
• Experimental results confirm the effectiveness and improvements of the SFDNet framework compared to state-of-the-art methods.

Conclusions:

• The novel SFDNet effectively addresses the limitations of existing VPR methods by incorporating structural information.
• The approach shows significant improvements in place recognition accuracy and robustness across diverse and challenging datasets.
• The findings suggest a promising direction for developing more reliable long-term visual place recognition systems.