Object detectors are typically learned based on fully-annotated training data with fixed pre-defined categories. However, not all possible categories of interest can be known beforehand, classes are often required to be increased progressively in many realistic applications. In such scenario, only the original training set annotated with the old classes and some new training data labeled with the new classes are available. Based on the limited datasets without extra manual labor, a unified detector that can handle all categories is strongly needed. Plain joint training leads to heavy biases and poor performance due to the incomplete annotations. To avoid such situation, we propose a practical framework in this paper. A conflict-free loss is designed to avoid label ambiguity, leading to an acceptable detector in one training round. To further improve performance, we propose a retraining phase in which Monte Carlo Dropout is employed to calculate the localization confidence, combined with the classification confidence, to mine more accurate bounding boxes, and an overlap-weighted method is employed for making better use of pseudo annotations during retraining to achieve more powerful detectors. Extensive experiments conducted on multiple datasets demonstrate the effectiveness of our framework for category-extended object detectors.
Download