Universal-RCNN: Universal Object Detector via Transferable Graph R-CNN.

摘要: The dominant object detection approaches treat each dataset separately and fit towards a specific domain, which cannot adapt to other domains without extensive retraining. In this paper, we address the problem of designing universal model that exploits diverse category granularity from multiple predict all kinds categories in one system. Existing works by integrating branches upon shared backbone network. However, paradigm overlooks crucial semantic correlations between domains, such as hierarchy, visual similarity, linguistic relationship. To these drawbacks, present novel detector called Universal-RCNN incorporates graph transfer learning for propagating relevant information across datasets reach coherency. Specifically, first generate global pool high-level representation categories. Then an Intra-Domain Reasoning Module learns propagates sparse within guided spatial-aware GCN. Finally, Inter-Domain Transfer is proposed exploit dependencies enhance regional feature attending transferring contexts globally. Extensive experiments demonstrate method significantly outperforms multiple-branch models achieves state-of-the-art results on benchmarks (mAP: 49.1% COCO).