It's official code implementation.
This study focuses on detecting ritual implements in Thangka paintings, an essential art form in Tibetan Buddhism. Due to their complex composition and small sizes, existing object detection models struggle in this domain. We propose Guided by Principles of Composition Detector (GPCDet), which incorporates Spatial Coordinate Attention (SCA) and Graph Convolution Network-Auxiliary Detection (GCN-AD) modules. The SCA module captures distinctive spatial features, while the GCN-AD learns co-occurrence relationships among categories. Extensive experiments on the Ritual Implements in Thangka (RITK) dataset demonstrate that GPCDet significantly outperforms previous methods, achieving an average precision of 38.2%, highlighting its effectiveness in accurately detecting ritual implements in Thangka art.
Please refer to mmdetection.