CamStyle: A Novel Data Augmentation Method for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (re-ID) is challenged by variations in image style across different cameras.
• Existing methods implicitly address this by learning camera-invariant features.

Purpose of the Study:

• To explicitly address camera style variations in person re-identification.
• To introduce a novel data augmentation technique, CamStyle, to improve re-ID model robustness and accuracy.

Main Methods:

• CamStyle utilizes style transfer to augment training data with diverse camera appearances.
• Label Smooth Regularization (LSR) is employed to mitigate noise introduced by style transfer.
• The method is evaluated on person re-identification benchmarks like Market-1501 and DukeMTMC-re-ID.

Main Results:

• CamStyle significantly reduces deep network overfitting and improves performance across various camera systems.
• The inclusion of LSR consistently enhances accuracy, especially in challenging scenarios.
• Competitive accuracy is achieved compared to state-of-the-art methods on standard datasets.

Conclusions:

• CamStyle effectively enhances person re-identification performance by addressing camera style variations.
• The approach shows particular promise for challenging tasks such as one-view learning and unsupervised domain adaptation (UDA).
• CamStyle achieves state-of-the-art results in UDA for person re-identification.