Inference Fusion with Associative Semantics for Unseen Object Detection

We study the problem of object detection when training and test objects are disjoint, i.e. no examples target classes available. Existing unseen approaches usually combine generic frameworks with a single-path classifier, by aligning regions semantic class embeddings. In this paper, inspired from human cognitive experience, we propose simple but effective dual-path model that further explores associative semantics to supplement basic visual-semantic knowledge transfer. use novel target-centric multiple-association strategy establish concept associations, ensure predictor generalized domain can be learned during training. way, through reasonable inference fusion mechanism, those two parallel reasoning paths strengthen correlation between seen objects, thus improving performance. Experiments show our inductive method significantly boost performance 7.42% over models, even 5.25% transductive models on MSCOCO dataset.