SKETCH BASED IMAGE RETRIEVAL (SBIR)

In this project, we look into the dilemma of zeroshot sketch-based image retrieval (ZS-SBIR), which involves using human sketches as queries to retrieve images from previously unseen categories. We make a significant contribution to prior art by proposing a ZS-SBIR scenario that represents a major advancement in its practical implementation. To exploit the domain void, highly abstract amateur human sketches are intentionally sourced rather than those found in existing datasets, which are mostly semi-photorealistic. The ZS-SBIR system is then used to model sketches and images into a single embedding space. To alleviate the domain gap, a novel strategy for mutual knowledge between domains has been created. To aid semantic transition, external semantic awareness is further embedded. Surprisingly, retrieval efficiency outperforms on existing datasets with a simplified version of our model. On the newly proposed dataset, we compare our complete model to a variety of alternatives and show that it outperforms them.