Towards a category-extended object detector with limited data

Object detectors are typically learned on fully-annotated training data with fixed predefined categories. However, categories often required to be increased progressively. Usually, only the original set annotated old classes and some new labeled available in such scenarios. Based limited datasets, a unified detector that can handle all is strongly needed. We propose practical scheme achieve it this work. A conflict-free loss designed avoid label ambiguity, leading an acceptable one round. To further improve performance, we retraining phase which Monte Carlo Dropout employed calculate localization confidence mine more accurate bounding boxes, overlap-weighted method proposed for making better use of pseudo annotations during retraining. Extensive experiments demonstrate effectiveness our method.