RD-OpenMax: Rethinking OpenMax for Robust Realistic Open-Set Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing open-set recognition (OSR) settings are often too idealized and fail to capture real-world complexities.
• Current OSR approaches struggle with scenarios like fine-grained classification, few-shot learning, long-tailed distributions, dynamic inputs, and cross-domain adaptation.

Purpose of the Study:

• To propose a realistic open-set recognition (ROSR) setting that encompasses diverse and challenging real-world scenarios.
• To introduce a novel method, regularized discriminative OpenMax (RD-OpenMax), designed to handle the complexities of the ROSR setting.

Main Methods:

• Developed a realistic OSR (ROSR) setting incorporating fine-grained, few-shot, long-tailed, dynamic input, and cross-domain adaptation challenges.
• Introduced RD-OpenMax, enhancing OpenMax with a covariance attention-based covariance pooling (CACP) module for discriminative distance scores.
• Proposed a regularized EVT (REVT) method using Monte Carlo sampling to stabilize extreme value theory estimation in few-shot and long-tailed scenarios.

Main Results:

• The proposed ROSR setting effectively challenges existing state-of-the-art OSR methods.
• RD-OpenMax significantly outperforms existing methods within the challenging ROSR setting.
• RD-OpenMax demonstrates competitive performance against state-of-the-art approaches in traditional OSR settings.

Conclusions:

• The novel RD-OpenMax method, utilizing CACP and REVT, provides a robust solution for open-set recognition in realistic, complex environments.
• The proposed ROSR setting serves as a more practical benchmark for evaluating OSR algorithms.
• RD-OpenMax offers a significant advancement in distinguishing known from unknown classes in diverse, real-world conditions.