Scene designer: compositional sketch-based image retrieval with contrastive learning and an auxiliary synthesis task

Leo Sampaio Ferraz Ribeiro, Tu Bui, John P. Collomosse, Moacir Ponti

Abstract: Scene Designer is a novel method for Compositional Sketch-based Image Retrieval (CSBIR) that combines semantic layout synthesis with its main task both to boost performance and enable new creative workflows. While most studies on sketch focus on single-object retrieval, we look to multi-object scenes instead for increased query specificity and flexibility. Our training protocol improves contrastive learning by synthesising harder negative samples and introduces a layout synthesis task that further improves the semantic scene representations. We show that our object-oriented graph neural network (GNN) more than doubles the current SoTA recall@1 on the SketchyCOCO CSBIR benchmark under our novel contrastive learning setting and combined search and synthesis tasks. Furthermore, we introduce the first large-scale sketched scene dataset and benchmark in QuickDrawCOCO.