Adaptive open domain recognition by coarse-to-fine prototype-based network

Highlights：

• For open domain recognition task, the overlap between known and unknown categorys, which is defined as openness, can affect the recognition model a lot. In this work, a realistic setting termed named as Adaptive Open Domain Recognition is firstly introduced to consider this issue. Fusion Information Guided Feature Prototype Generation is proposed to obtain more abundant and accurate descriptive information.

• FGPG is more robust to various openness since it directly constructs the interaction between visual features and fused category semantics through generative adversarial networks.

• Class-Aware Feature Prototype Alignment is proposed to align the global feature prototype between two domains. The global feature prototype is adaptively updated in each episode to suppress the negative effects of false pseudo-label and misaligned categories, and further adapt to various openness.