Dynamic Correlation Learning and Regularization for Multi-Label Confidence Calibration.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Modern visual recognition models exhibit overconfidence, leading to unreliable confidence scores due to deep neural networks and one-hot encoding.
• Existing confidence calibration methods primarily focus on single-label scenarios, neglecting the complexities of multi-label image recognition.
• Multi-label images present unique challenges, including semantic confusion and unreliable confidence scores, due to the presence of multiple objects.

Purpose of the Study:

• To introduce the Multi-Label Confidence Calibration (MLCC) task for generating well-calibrated confidence scores in multi-label visual recognition.
• To address the limitations of existing single-label calibration methods that fail to account for crucial category correlations in multi-label contexts.
• To propose a novel algorithm, Dynamic Correlation Learning and Regularization (DCLR), for adaptive regularization in multi-label scenarios.

Main Methods:

• Developed the Dynamic Correlation Learning and Regularization (DCLR) algorithm to leverage multi-grained semantic correlations for modeling semantic confusion.
• DCLR learns dynamic instance-level and prototype-level similarities to measure inter-category semantic correlations.
• Constructed adaptive label vectors based on learned correlations to facilitate more effective regularization.

Main Results:

• Established an evaluation benchmark by re-implementing and applying advanced confidence calibration algorithms to leading multi-label recognition (MLR) models.
• Extensive experiments demonstrated that DCLR significantly outperforms existing methods in providing reliable confidence scores for multi-label scenarios.
• The proposed DCLR algorithm effectively mitigates semantic confusion, leading to improved calibration performance.

Conclusions:

• The Multi-Label Confidence Calibration (MLCC) task and the DCLR algorithm provide a robust solution for improving confidence reliability in multi-label visual recognition.
• DCLR's ability to model dynamic category correlations offers a significant advancement over traditional label smoothing techniques.
• The findings highlight the importance of considering semantic correlations for effective confidence calibration in complex multi-label visual recognition tasks.