Image slicing creates some wrong segmentations

[s-fx]

Hi,
I'm working on a ML segmentation project with high resolution images. I sliced the images into small patches of 256x256 Pixels and the COCO format Json file respectively (containing bboxes and segmentations).

coco_dict, coco_path = slice_coco(
         coco_annotation_file_path=json_file_dir,
         image_dir=img_dir,
         output_coco_annotation_file_name=root_dir + "sliced_annotations",
         output_dir=root_dir + dir_,
         slice_height=img_height,
         slice_width=img_width,
         overlap_height_ratio=0.2,
         overlap_width_ratio=0.2,
         min_area_ratio=0.1,
         verbose=False,
     )

But when I visualize the end results (with fiftyone) some of my segmentations are now just "filled bounding boxes" and some of the segmentations are correctly sliced. Unfortunately I can not share the images, but I am wondering why some of my segmentations are now just in the shape of a bounding box.

While slicing I encounter this Error message:

ERROR - shapely.geos -   TopologyException: Input geom 0 is invalid: Ring Self-intersection at 3201 1619
09/02/2022 15:50:43 - INFO - shapely.geos -   Ring Self-intersection at or near point 3201 1619

Could this be the cause because of polygons intersecting?

=> sahi 0.10.4

kind regards

[fcakyon]

Hello @s-fx, the shapely.geos error you posted means there are self intersections in your polygons after slicing (or before slicing). Have you visualized the polyons begore slicing operation?

[s-fx]

Yes I have visualized them before slicing, the segmentation were visualized correctly (no filled bounding boxes as segmenations), but some look like they could have intersecting polygones. So after creating patches these "wrong" segmentations come from intersecting polygones?