Answers are grounded in this paper and cite their sources.
Autonomous vehicles rely on accurate traffic sign classification, which is typically achieved through supervised learning. However, the diversity and complexity of traffic signs make it impractical to rely solely on large labeled datasets. While abundant data exists for common signs such as stop and yield signs, less common signs often lack sufficient representation in existing datasets. Few-shot learning has been proposed as an alternative solution for such cases in which there is not enough training data, but its effectiveness decreases as the number of classes increases. To address these challenges, our research introduces an innovative adaptive hierarchical framework with contrastive aggregation (HF-CA). This framework strategically reduces class dimensionality and enriches the dataset with more examples per category through contrastive aggregation. We validated our approach using modified versions of the GTSRB and Mapillary datasets, demonstrating that our method consistently outperforms existing baselines. By simplifying the classification process, our solution enhances classification accuracy and provides a scalable approach for scenarios with numerous classes but limited labels.
No takes yet. Share an insight, caveat, or question.
Valiente et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: