PyrAug: Gaussian Pyramid-based Multi-scale Data Augmentation for Panoptic Plant Segmentation with Ambiguous Boundary

Most of the plants are usually roughly assigned categories in previous panoptic segmentation, which cannot meet the demand for autonomous mobile robotic sprayer sys-tems. Only a reasonable allocation of pesticide doses to different kinds of plants can keep them from being harmed by excessive chemicals. Due to the ambiguous boundaries and similar feature distributions between different plants, implementing precise plant segmentation is still a challenging task. To address this challenge, traditional data augmentation methods enhance the feature extraction capability of the model with geometric and color enhancements. However, these methods only augment the data at the image level such that the implicit scale information in the image is ignored. In this paper, we propose a Gaussian pyramid - based multi -scale data augmentation for panoptic segmentation of plants with ambiguous boundaries, named PyrAug. Our methods attempt to hierarchically enhance images at different levels, in which the network can effectively learn different feature distributions among plants and thus predict the respective masks. In addition to this, we present a new panoptic plant segmentation dataset for improving the seg-mentation performance in complex landscapes. In experimental verification, our method is added to the advanced panoptic segmentation network, and results show that our methods outperform the baselines with higher panoptic quality scores.