BDA-SketRet: Bi-level domain adaptation for zero-shot SBIR

The efficacy of zero-shot sketch-based image retrieval (ZS-SBIR) models is governed by two challenges. immense distributions-gap between the sketches and images requires a proper domain alignment. Moreover, fine-grained nature task high intra-class variance many categories necessitates class-wise discriminative mapping among sketch, image, semantic spaces. Under this premise, we propose BDA-SketRet, novel ZS-SBIR framework performing bi-level adaptation for aligning spatial features visual data pairs progressively. In order to highlight shared reduce effects any sketch or image-specific artifacts, symmetric loss function based on notion information bottleneck while cross-entropy-based adversarial introduced align feature maps. Finally, our CNN-based model confirms discriminativeness latent space through topology-preserving projection network. Experimental results extended Sketchy, TU-Berlin, QuickDraw datasets exhibit sharp improvements over literature.