Amplitude Spectrum Transformation for Open Compound Domain Adaptive Semantic Segmentation

Open compound domain adaptation (OCDA) has emerged as a practical setting which considers single labeled source against of multi-modal unlabeled target data in order to generalize better on novel unseen domains. We hypothesize that an improved disentanglement domain-related and task-related factors dense intermediate layer features can greatly aid OCDA. Prior-arts attempt this indirectly by employing adversarial discriminators the spatial CNN output. However, we find latent derived from Fourier-based amplitude spectrum deep hold more tractable mapping with discrimination. Motivated this, propose feature space Amplitude Spectrum Transformation (AST). During adaptation, employ AST auto-encoder for two purposes. First, carefully mined source-target instance pairs undergo simulation cross-domain stylization (AST-Sim) at particular altering AST-latent. Second, operating later is tasked normalize (AST-Norm) content fixing its mean prototype. Our simplified technique not only clustering-free but also free complex alignment. achieve leading performance prior arts OCDA scene segmentation benchmarks.