Weakly Supervised Object Detection Based on Active Learning

Weakly supervised object detection which reduces the need for strong supersivison during training has recently made significant achievements. However, it remains a challenging issue due to the time-consuming and labor-intensive problems in application. To further reduce the label cost, we introduce a new fusion method of weakly supervised learning and active learning in a unied framework for object detection. Weakly supervised learning based on min-entropy latent model is used to weaken the labels by image-label, while active learning is used to reduce the quantity of labeled images. The fusion method proposed can effectively reduce the dependency of object detection on manual annotation. In this paper, we introduce three strategies of active learning, including least confidence sampling, margining sampling and weighted classification sampling. To validate the effectiveness of each strategy and different sample compositions in weakly supervised learning object detection, we conducted lots of experiments. Extensive experiments show that the combination of image-level labeling and active learning can achieve comparable results with the previous state-of-the-art methods with much lower label cost.