Hierarchical Novelty Detection for Traffic Sign Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Novelty detection identifies unknown samples but offers limited information.
• Existing methods lack informative outputs for novel class instances.

Purpose of the Study:

• To develop a hierarchical novelty detection method for informative novel class identification.
• To predict the parent class within a taxonomy for novel samples.

Main Methods:

• Proposed a novel Hierarchical Cosine Loss function.
• Learned class prototypes and discriminative features aligned with class taxonomy.
• Applied the method to traffic sign recognition and natural image datasets.

Main Results:

• Achieved state-of-the-art performance on traffic sign benchmarks (MTSD, TT100K).
• Demonstrated high accuracy in detecting novel samples at correct hierarchical nodes (81% on TT100K, 36% on MTSD).
• Maintained competitive performance on natural image datasets (AWA2, CUB).

Conclusions:

• Hierarchical novelty detection offers informative outputs for unknown samples.
• The Hierarchical Cosine Loss effectively learns class prototypes and discriminative embeddings.
• The proposed method significantly advances novelty detection, particularly in specialized domains like traffic sign recognition.