CGAT-Net: Context-Aware Graph Attention Transformer Network for Scene Sketch Recognition

Sketches often lack sufficient detail or quality necessary for standalone recognition, making their identification challenging without contextual information. While context understanding is commonly studied in computer vision applications like object detection or image classification, it remains under-explored in the sketch domain. Existing research primarily focuses on recognizing sketch objects in isolation, with little attention given to scene-level sketch understanding. To address this gap, we introduce a Context-Aware Graph Attention Transformer Network (CGAT-Net), which leverages visual and spatial relationships among objects to obtain a more accurate classification within a scene. This is the first study in scene sketch recognition that utilizes object relations in a Transformer-based network to incorporate context understanding. Extensive experiments show that CGAT-Net surpasses current state-of-the-art single-sketch classifiers, underscoring the value of contextual information in enhancing individual sketch recognition. Our code and trained model weights can be accessed from https://github.com/aleynakutuk6/CGAT-Net.