This paper describes a novel method of training a semantic segmentation model for scene recognition of agricultural mobile robots exploiting multiple publicly available datasets that are different from the target greenhouse environments. Semantic segmentation models require abundant labels given by tedious manual annotation for training. Although unsupervised domain adaptation (UDA) is studied as a workaround for such a problem, existing UDA methods assume a source dataset similar to the target dataset, which is not available for greenhouse scenes. In this paper, we propose a method to train a semantic segmentation model for greenhouse images leveraging multiple publicly available datasets not dedicated to greenhouses. We exploit segmentation models pre-trained on each source dataset to generate pseudo-labels for the target images based on agreement of all the pre-trained models on each pixel. The proposed method allows for effectively transferring the knowledge from multiple sources rather than relying on a single dataset and realizes precise training of semantic segmentation model. We also introduce existing state-of-the-art methods to suppress the effect of noise in the pseudo-labels to further improve the performance. We demonstrate that our proposed method outperforms existing UDA methods and a supervised SVM-based method.
展开▼
