Mask2KAN: A Universal Image Segmentation Kolmogorov–Arnold Network Architecture

Universal architectures like Mask2Former have redefined the way we approach image segmentation tasks. Traditionally, specialized architectures were used for specific tasks such as semantic, instance, and panoptic segmentation. Now, a single, unified architecture can outperform these task-specific models, offering benefits in performance, efficiency, and effort, while also reshaping the way we perceive these tasks. In this paper, experiments are conducted using the Mask2Former configuration for \textit{semantic segmentation}. However, similar to other universal models like DETR, these architectures, despite sharing the same underlying structure, \textit{are still trained separately for different tasks and datasets}. Recent works on the passage from the Universal Approximation Theorem to a Kolmogorov-Arnold theorem inspired the present work to delve in Kolmogorov Arnold Network on computer vision tasks. Traditional semantic segmentation models as Mask2Former, recognize a predefined set of classes, often failing to detect unseen objects (anomalies). To address this, we propose Mask2KAN, a novel approach derived from the Mask2Former architecture , which shifts from a per-pixel (i.e. BERT) to a mask classification (i.e. Mask2Former) focusing on reducing ood anomalies (i.e. Mask2Anomaly), with an efficient Kolmogorov-Arnold Network (KAN) mask embed prediction head, hence improving the segmentation of unseen objects and reducing false positives. Proposed architectures include ResNet-50 and Swin-T/S/B/L as backbones. and using KAN mask embed layers sets a new state-of-the-art in anomaly segmentation, since our approach demonstrates superior performance across various benchmarks on semantic segmentation, making it a robust solution also for real-world scenarios as autonomous driving applications or anomaly detection in the wild. For more details and code, visit our \href{https://github.com/gguzzy/benchmark}{Github page}.