Utilizing Google Images for Training Classifiers in CRF-Based Semantic Segmentation

抄録

One promising approach to pixel-wise semantic segmentation is based on conditional random fields (CRFs). CRF-based semantic segmentation requires ground-truth annotations to supervisedly train the classifier that generates unary potentials. However, the number of (public) annotation data for training is limitedly small. We observe that the Internet can provide relevant images for any given keywords. Our idea is to convert keyword-related images to pixel-wise annotated images, then use them as training data. In particular, we rely on saliency filters to identify the salient object (foreground) of a retrieved image, which mostly agrees with the given keyword. We utilize saliency information for back-and-foreground CRF-based semantic segmentation to further obtain pixel-wise ground-truth annotations. Experiment results show that training data from Google images improves both the learning performance and the accuracy of semantic segmentation. This suggests that our proposed method is promising for harvesting substantial training data from the Internet for training the classifier in CRF-based semantic segmentation.