Body segmentation is a process of labelling parts of images and classifying them into semantic classes of body parts and clothing pieces. Body segmentation is used in many complex systems that process images of people and is one of the key parts of such processing pipelines. In this work, we try to improve the existing models for human segmentation using multi-task learning method. We present and implement a multi-task model SPD, which, in addition to segmentation, also includes the tasks of predicting the body skeleton and the prediction of dense pose. We also implement other models that include only one additional task next to the segmentation task. We prepare a data set of images that contain all three necessary annotations types for learning. The results of the implemented models are analysed and compared with the results of JPPNet reference model and DensePose model. The Segmentation results indicate an improvement in all metrics in comparison to the JPPNet segmentation model. The results of skeletal and dense pose representations perform a little worse than the reference models.
|
