Recurrent Instance Segmentation

摘要: Instance segmentation is the problem of detecting and delineating each distinct object interest appearing in an image. Current instance approaches consist ensembles modules that are trained independently other, thus missing opportunities for joint learning. Here we propose a new paradigm consisting end-to-end method learns how to segment instances sequentially. The model based on recurrent neural network sequentially finds objects their segmentations one at time. This net provided with spatial memory keeps track what pixels have been explained allows occlusion handling. In order train designed principled loss function accurately represents properties problem. experiments carried out, found our outperforms recent multiple person segmentation, all state art Plant Phenotyping dataset leaf counting.