Multi-Domain Semantic-Segmentation using Multi-Head Model

Semantic segmentation is a pixel-wise class identification problem, which is important for automatic driving support such as recognizing the driving area. However, segmentation accuracy significantly degrades in scenes that differ from the training domain. Therefore, it is necessary to prepare multiple models for each domain, which increases the memory cost. When training multiple datasets with a single-head model, it is also necessary to redefine a different object class for each dataset. We propose a semantic-segmentation method that involves using a multi-head model for supporting multiple domains. The proposed method also involves using a shared network for sharing all domains for training datasets. This makes it possible to train multiple datasets with different object classes in a single network. To train all datasets equally, we also introduce mix loss, which simultaneously back-propagates the loss of each dataset. From experiments evaluating the proposed method, we confirmed that the method achieves higher or equivalent recognition accuracy with fewer parameters than using a single-head model for each dataset when training datasets with the same class, training different datasets at the same time, and training datasets individually.