Unsupervised Cross Domain Person Re-Identification by Multi-Loss Optimization Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised cross-domain (UCD) person re-identification (re-ID) is challenging due to non-overlapping identities between source and target domains.
• Existing UCD person re-ID methods often rely on noisy pseudo-labels from supervised learning, leading to suboptimal performance.
• Pseudo-label noise significantly hinders the effectiveness of current UCD person re-ID approaches.

Purpose of the Study:

• To propose a novel multi-loss optimization learning (MLOL) model for UCD person re-ID.
• To address the limitations of pseudo-label noise in existing UCD person re-ID methods.
• To enhance the accuracy and robustness of person re-identification across different domains without labeled target data.

Main Methods:

• Developed a multi-loss optimization learning (MLOL) model incorporating supervised learning with clustering pseudo-labels.
• Introduced two novel losses: ranking-average-based triplet loss for similarity exploration and neighbor-consistency-based loss for adversarial learning.
• Designed losses to mitigate errors from clustering and deeply explore intra-domain relations within the target domain.

Main Results:

• The proposed MLOL model demonstrated superior performance on three benchmark datasets: Market-1501, DukeMTMC-reID, and MSMT17.
• Experimental results show a clear advantage over state-of-the-art UCD re-ID methods.
• The model effectively reduced the negative impact of pseudo-label noise on re-identification accuracy.

Conclusions:

• The MLOL model offers a significant advancement in UCD person re-identification by effectively handling pseudo-label noise.
• The combination of similarity exploration and adversarial learning losses enhances the model's ability to explore target domain relations.
• This approach provides a more robust and accurate solution for person re-identification in unsupervised cross-domain scenarios.