earticle

한국어

This paper presents an improved approach to generate pseudo labels for unlabeled dataset. To properly train a network, large amount of dataset is required. The publicly available datasets are often not large enough or versatile. Although we can acquire a great deal of images from the internet, those images are not labeled. Conventionally, the generation of ground truth labels requires human effort which is very expensive and time-consuming. Recently, existing object detectors are being employed to automate the generation of labels, called pseudo labels. Such pseudo labels have poor accuracy, since most of the object detectors employ simplistic confidence thresholding, which tends to discard even good labels. This paper proposes an enhanced pseudo labeling technique that selects the predicted labels using a bi-directional tracking method instead of simplistic confidence thresholding. The proposed technique can recover many predicted labels that are actual good labels but would have been discarded due to their poor confidence. Our method can produce pseudo labels for new training dataset with higher accuracy than conventional pseudo labeling techniques, thus offering better training accuracy for object detector CNN models.