Drop Loss for Person Attribute Recognition With Imbalanced Noisy-Labeled Samples.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning methods for Person Attribute Recognition (PAR) show strong performance.
• Existing PAR models struggle with imbalanced noisy-labeled data, impacting overall accuracy.
• Attribute imbalance and noise in training datasets lead to suboptimal PAR performance.

Purpose of the Study:

• To propose a novel loss function, Drop Loss, to address imbalanced noisy-labeled samples in PAR.
• To improve the robustness and accuracy of person attribute recognition models.
• To mitigate the negative effects of noisy and imbalanced data on deep learning models.

Main Methods:

• Introduced Drop Loss, a loss function that treats attributes differently based on difficulty (easy-to-hard).
• Identified noisy-labeled candidates using gradient norms and applied a higher drop rate for harder attributes.
• Developed DropNet, a ResNet-50 model incorporating the proposed Drop Loss for PAR tasks.

Main Results:

• DropNet achieved comparable or superior performance compared to state-of-the-art PAR methods.
• The proposed Drop Loss effectively alleviated the adverse effects of imbalanced noisy-labeled samples.
• Experiments on facial and pedestrian attribute recognition tasks validated the method's effectiveness.

Conclusions:

• Drop Loss is an effective strategy for handling imbalanced noisy labels in person attribute recognition.
• The proposed DropNet model demonstrates significant improvements in balanced and classification accuracy.
• This approach offers a promising direction for enhancing deep learning-based attribute recognition systems.