Cross-Domain Pedestrian Attribute Recognition: Evaluation Criteria, a New Baseline and Remote Sensor-Based...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pedestrian attribute recognition (PAR) is crucial for tasks like person re-identification.
• Domain differences between datasets cause significant performance degradation in mainstream PAR methods.
• This issue is amplified in remote sensing PAR, where models trained on fixed-camera data struggle with UAV-based data.

Purpose of the Study:

• To formally introduce the task of cross-domain pedestrian attribute recognition (CD-PAR).
• To establish evaluation criteria for the new CD-PAR task.
• To propose an effective baseline method for CD-PAR.

Main Methods:

• Introduced the novel task of cross-domain pedestrian attribute recognition (CD-PAR).
• Developed a new baseline method, local domain discriminator-based cross-domain pedestrian attribute recognition (LDCD-PAR).
• Employed adversarial training with a local domain discriminator to achieve fine-grained domain-invariant features.

Main Results:

• Demonstrated the value of the CD-PAR task through extensive cross-domain experiments.
• Validated the effectiveness of the proposed LDCD-PAR method on remote sensor-based PAR.
• Showcased significant improvements in cross-domain performance.

Conclusions:

• The proposed CD-PAR task is essential for advancing PAR in diverse, real-world scenarios.
• LDCD-PAR effectively addresses domain shift challenges in pedestrian attribute recognition.
• The findings pave the way for more robust and adaptable PAR systems, particularly in remote sensing applications.