Soft Transfer Learning via Gradient Diagnosis for Visual Relationship Detection

摘要: Detecting all visual relationships is posed as the most fundamental task towards ultimate semantic reasoning. However, due to rich context embedded in image and diverse language ambiguities, it unrealistic annotate list possible for providing a noise-free supervised setting. All prior approaches simply adopt traditional fully-supervised detection pipeline ignore effect of incomplete annotations on model convergence, resulting unstable optimization unsatisfactory performance. In this work, we make first attempt address critical issue reformulate via Soft Transfer Learning (STL), which aims transfer knowledge learned from hand into uncertain pairs self-supervised way. The process inferred principled gradient diagnosis. Extensive experiments VRD large-scale VG benchmarks demonstrate superiority our STL method.