Zero-Shot Detection with Transferable Object Proposal Mechanism

Abstract

Zero-shot detection (ZSD) aims to locate and recognize unseen objects in pictures or videos to address the shortage of labeled training data. While most existing ZSD methods lay their emphasis on object classification, challenges lie in both object proposal and category prediction for detectors to get over domain shift. In this paper, we first design an experiment to verify the impact of transfer ability of the object proposal step on detection recall and further introduce a transferable mechanism to relate the co-occurrence among categories. We use a confidence distribution over all the classes for object confidence prediction. Experimental results show our method outperforms other zero-shot detectors on PASCAL VOC and MSCOCO datasets, even with a simple linear approach for classification.