Multibox Sample Selection for Active Object Detection.

Deep object detection has achieved state-of-the-art performance but highly relies on large-scale labeled datasets. Active Object Detection (AOD) presents criteria to select the most informative samples to annotate, thereby maximizing the performance with limited labeled data. But current AOD methods mostly aggregate criterion values calculated from every single box without the MultiBox prediction structure of object detection, making them easily overwhelmed by uninformative boxes (e.g. boundary and background boxes). This paper presents a MultiBox Sample Selection (MSS) criterion composed of MultiBox Uncertainty (MBU) and MultiBox Committee (MBC). MBU eliminates the influence of uninformative boxes and selects the boxes that best reflect the prediction uncertainty of the whole image. MBC consists of a committee measurement, which is based on the fact that certain but incorrect box predictions are usually inconsistent among nearby boxes. Experimental results on in-domain and cross-domain object detection datasets show the highest performance under the same labeling budget.