In the roof_fine_train.json file, there are three fields: 'images', 'annotations', and 'categories'. The annotations field includes several subfields, such as 'segmentation', 'iscrowd', 'image_id', 'bbox', 'area', and 'id'. The segmentation field is supposed to contain a list of polygon corner points, but for some examples, it is an object with 'size' and 'count' fields. Is this an error or is it intended?
Posted by: duraklefkan @ Jan. 4, 2023, 7:40 p.m.The 'segmentation' field contains two kinds of mask format. For format with 'size' and 'count' fields, it's an run-length encoding (RLE). The main reason for using the RLE format is to enable the representation of masks for buildings with interior cavity structures. If you are using python, you can use pycocotools to decode RLE objects.
Posted by: dfc2023.iecas @ Jan. 5, 2023, 3:42 a.m.