ZeroDivisionError from Empty MultiPolygon in slicing.process_coco_annotations

[matheuslima29]

Description

An error ZeroDivisionError: division by zero is thrown when running slice_coco function, specifically after handling GeometryCollection objects that do not properly account for nested MultiPolygon structures, leading to an empty MultiPolygon with an area of zero.

The exception is thrown at:

sahi/sahi/slicing.py

Line 145 in 16c45ac

if sliced_coco_annotation.area / coco_annotation.area >= min_area_ratio:

The problem arises specifically within the filter_polygons function, which fails to correctly process nested MultiPolygon objects inside GeometryCollection. Check out the full function here:

sahi/sahi/utils/coco.py

Lines 238 to 240 in 16c45ac

	elif isinstance(geometry, GeometryCollection):
	polygons = [geom for geom in geometry.geoms if isinstance(geom, Polygon)]
	return MultiPolygon(polygons) if polygons else MultiPolygon()

Steps to Reproduce

1. Configure the slice_coco function with a typical set-up using the provided JSON dataset.
2. Execute the function to process annotations which might include GeometryCollection objects.
3. Observe the ZeroDivisionError when the area of a resulting MultiPolygon is zero due to improper filtering.

Environment Setup

• Operating System: Windows 11
• Python Version: 3.11 (Anaconda Environment)
• Environment: Isolated Anaconda environment with only project dependencies installed from pip.

You can install these dependencies using the following pip command:

pip install sahi

Code Exemple

from sahi.slicing import slice_coco

coco_annotation_file_path = 'path/to/coco_segmentation.json'
image_dir = 'path/to/images'
output_coco_annotation_file_name = './coco_sahi_segmentation.json'
output_dir='./path/to/output'

size = 320
ratio = 0.2

coco_dict, coco_path = slice_coco(
    coco_annotation_file_path=coco_annotation_file_path,
    image_dir=image_dir,
    output_coco_annotation_file_name=output_coco_annotation_file_name,
    slice_height=size,
    slice_width=size,
    overlap_height_ratio=ratio,
    overlap_width_ratio=ratio,
    output_dir=output_dir,
)

Sample coco_segmentation.json

{
  "images": [
    {
      "file_name": "Image1.png",
      "height": 3507,
      "width": 2480,
      "id": 1
    }
  ],
  "annotations": [
    {
      "id": 488,
      "image_id": 1,
      "category_id": 20,
      "bbox": [689, 2000, 78, 98],
      "area": 7644,
      "iscrowd": 0,
      "segmentation": [
         [
           709, 2000, 705, 2005, 705, 2038, 728, 2038, 740, 2054, 740, 2065, 
           736, 2071, 736, 2076, 732, 2082, 701, 2082, 697, 2076, 689, 2076, 
           693, 2082, 693, 2087, 701, 2098, 720, 2098, 740, 2071, 740, 2049, 
           744, 2043, 751, 2043, 755, 2038, 759, 2038, 763, 2032, 763, 2027, 
           767, 2022, 767, 2011, 763, 2011, 759, 2005, 755, 2005, 751, 2000
         ]
      ]
    }
  ],
  "categories": [
    {
      "id": 1,
      "name": "Object",
      "supercategory": "none"
    }
  ]
}

Expected Behavior

The function should handle nested MultiPolygon objects within GeometryCollection without returning an empty MultiPolygon, thereby avoiding division by zero when calculating area ratios.

Actual Behavior

The filter_polygons function fails to correctly process MultiPolygon objects nested within GeometryCollection. This results in an empty MultiPolygon being returned, and when the area of this MultiPolygon is accessed, a ZeroDivisionError occurs due to the area being zero.

Suggested Fix:

We should tweak the filter_polygons method to ensure it captures all Polygon and MultiPolygon structures within GeometryCollection.

elif isinstance(geometry, GeometryCollection):
    polygons = []
    for geom in geometry.geoms:
        if isinstance(geom, Polygon):
            polygons.append(geom)
        elif isinstance(geom, MultiPolygon):
            polygons.extend([subgeom for subgeom in geom.geoms if isinstance(subgeom, Polygon)])

This fix addresses the issue where nested polygons exist but still leaves open the possibility of an empty MultiPolygon resulting from other scenarios.

sahi/sahi/utils/coco.py

Lines 240 to 241 in 16c45ac

	return MultiPolygon(polygons) if polygons else MultiPolygon()
	return MultiPolygon()

Handling Remaining Issue:

I can open a PR for the suggested fix, but we need a strategy for handling the empty MultiPolygon scenario. My suggestion is to skip the annotation and log a warning whenever this happens:

if sliced_coco_annotation.is_empty:
    logger.warning(f"Skipping annotation of image_id {coco_annotation.image_id}, due to an empty polygon")
    continue

I plan to implement the .is_empty method on both CocoAnnotation and ShapelyAnnotation based on Shapely's native MultiPolygon().is_empty method to ensure robust checks and align with the project's code structure.

[fcakyon]

Thanks for the detailed analysis, @matheuslima29 ! It would be amazing if you could open a PR with your suggestions 💯