SIF: Self-Inspirited Feature Learning for Person Re-identification

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Person re-identification (ReID) research has focused on new network architectures for improved person representation learning.
• Existing ReID networks are typically trained with standard optimizers, with limited exploration into optimizing training schemes for enhanced performance.

Purpose of the Study:

• To propose a novel training method, Self-Inspirited Feature Learning (SIF), to boost the performance of existing ReID networks.
• To enhance person representation learning through an adversarial training scheme that encourages more discriminative features.

Main Methods:

• Introduced Self-Inspirited Feature Learning (SIF), an optimization-focused method for ReID.
• Implemented a simple adversarial learning scheme with an auxiliary branch, active only during training.
• Ensured the original network architecture remains unchanged during testing.

Main Results:

• SIF demonstrated significant performance improvements across three public ReID datasets: Market1501, DuckMTMC-reID, and CUHK03.
• Achieved state-of-the-art results, including 87.6% mAP / 95.2% Rank-1 on Market1501 and 77.0% mAP / 79.5% Rank-1 on CUHK03 (labeled).
• SIF is general, compatible with various ReID networks, easy to implement, and provides stable performance.

Conclusions:

• Self-Inspirited Feature Learning (SIF) effectively enhances existing person re-identification networks through improved optimization.
• The proposed adversarial training scheme leads to more discriminative feature learning, achieving state-of-the-art results.
• SIF offers a versatile and practical approach to advancing person re-identification performance.