Rewarded Semi-Supervised Re-Identification on Identities Rarely Crossing Camera Views

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised person re-identification (Re-ID) methods often assume abundant cross-camera identities, which is unrealistic in many real-world applications.
• This assumption fails in scenarios with non-adjacent camera views, leading to uncertain sample relations and noise accumulation in existing Re-ID techniques.

Purpose of the Study:

• To develop a semi-supervised Re-ID method that addresses the challenge of identities rarely crossing camera views.
• To mitigate the noise accumulation problem caused by uncertain sample relations in existing pseudo-labeling approaches.

Main Methods:

• Introduced Rewarded Relation Discovery (R²D) to parameterize probabilistic sample relations for pseudo-label training.
• Utilized a reward mechanism based on identification performance on labeled data to guide dynamic relation learning and reduce uncertainty.
• Employed multiple relation discovery objectives and similarity distillation to fuse complementary knowledge from intra-camera affinity and cross-camera style variations.

Main Results:

• The proposed R²D method significantly outperforms existing semi-supervised and unsupervised Re-ID methods.
• Demonstrated effectiveness on a newly collected dataset (REID-CBD) and benchmark datasets under the condition of rare cross-camera identities.
• Successfully reduced uncertainty in sample relations, a key challenge in Re-ID with limited cross-camera data.

Conclusions:

• Rewarded Relation Discovery (R²D) offers a novel and effective approach for semi-supervised person Re-ID in challenging real-world settings.
• The rewarded learning paradigm and fusion of probabilistic relations are crucial for handling uncertain sample associations.
• The study highlights the importance of addressing the rare cross-camera identity scenario for practical Re-ID applications.