The Fieldscapes Dataset for Semantic Field Scene Understanding

Image segmentation is the focus of research in the field of computer vision, and it is the perceptual basis of robots and autonomous vehicles. The rapid development of image segmentation is greatly attributed to the combination of large-scale datasets and deep learning related methods. But to our knowledge for the understanding of semantic field scene, there are currently very few datasets that can reflect the complexity of real-world field environments. To address this, we introduce the Fieldscapes dataset, which contains pixel-level and instance-level semantic labeling. The objects and backgrounds appearing in the dataset are divided into 12 categories, including 7 types of instance-level annotations. The images captured in the fields through the front view of the autonomous vehicle provide more information for its understanding of the agricultural environments. In the meanwhile, in order to adapt to the diversity of the field environment, the images in the dataset are collected from different scenes, which ensures that the model generated by the neural network can identify various obstacles, vegetation and crops in the farmland to a certain extent. Fieldscapes has been evaluated by using different neural networks and further demonstrated the contribution of our dataset to the semantic understanding of field scenes.