A meaningful learning method for zero-shot semantic segmentation

Zero-shot semantic segmentation, which is developed to segment unseen categories, has attracted increasing attention due to its strong practicability. Previous approaches usually applied semantic-visual mapping based on seen categories to unseen categories, and thus failed to generate meaningful unseen visual representations and struggled to balance the seen and unseen concepts in the classifier. To overcome the above limitations, we propose a novel meaningful learning method that could be embedded into any generation-based zero-shot semantic segmentation model, borrowing the idea from the educational psychology field. The proposed meaningful learning method refers to the process that the new concepts could be learned by relating to existing comprehensible concepts and harmoniously incorporated into the concept schema. Specifically, we introduce a generator with conjugate conceptual correlation (G3C) which generates meaningful unseen visual information through anchoring into existing concepts. Moreover, simulating the rational thinking mechanism, we introduce a fast-slow concept modulator to alleviate the noisy over-correlation problem introduced by G3C and further construct a comprehensive concept schema. Extensive experiments conducted on three benchmarks demonstrate the superior performance of our method, especially according to the commonly-acknowledged h-mIoU (e.g., 4% improvement on the Pascal-VOC dataset).