Unsupervised Domain Adaptation on Person Reidentification Via Dual-Level Asymmetric Mutual Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation (UDA) person reidentification (Re-ID) addresses identifying pedestrians across different datasets without target domain labels.
• Existing methods often use homogeneous networks, limiting knowledge diversity and leading to shared errors.
• This limitation hinders the ability to learn robust and discriminative representations.

Purpose of the Study:

• To propose a novel dual-level asymmetric mutual learning (DAML) method for improved UDA person Re-ID.
• To enable learning from a broader knowledge scope through diverse embedding spaces.
• To overcome the limitations of homogeneous networks in existing approaches.

Main Methods:

• DAML utilizes two heterogeneous networks for mutual knowledge learning.
• Asymmetric subspaces and hard distillation are employed for pseudo-label generation.
• Knowledge transfer occurs via an asymmetric mutual learning (AML) strategy, with a teacher network adapting to the target domain and a student network learning from ground-truth labels.

Main Results:

• The proposed DAML method demonstrated superior performance compared to state-of-the-art methods.
• Experiments were conducted on challenging public datasets: Market-1501, CUHK-SYSU, and MSMT17.
• The results validate the effectiveness of learning from diverse embedding spaces and asymmetric knowledge transfer.

Conclusions:

• DAML effectively learns discriminative representations for UDA person Re-ID.
• The asymmetric mutual learning approach enhances model robustness by leveraging diverse knowledge.
• The method offers a significant advancement in cross-domain pedestrian identification.