WAD-CMSN: Wasserstein distance-based cross-modal semantic network for zero-shot sketch-based image retrieval

Zero-shot sketch-based image retrieval (ZSSBIR) aims at retrieving natural images given free hand-drawn sketches that may not appear during training. Previous approaches used semantic aligned sketch-image pairs or utilized memory expensive fusion layer for projecting the visual information to a low-dimensional subspace, which ignores significant heterogeneous cross-domain discrepancy between highly abstract sketch and relevant image. This yield poor performance in training phase. To tackle this issue overcome drawback, we propose Wasserstein distance-based cross-modal network (WAD-CMSN) ZSSBIR. Specifically, it first projects of each branch (sketch, image) common subspace via distance an adversarial manner. Furthermore, novel identity matching loss is employed select useful features, can only capture complete knowledge, but also alleviate over-fitting phenomenon caused by WAD-CMSN model. Experimental results on challenging Sketchy (Extended) TU-Berlin datasets indicate effectiveness proposed model over several competitors.