Participants' solutions are evaluated using the [mAP@50](https://learnopencv.com/mean-average-precision-map-object-detection-model-evaluation-metric/) metric. The metric code is available [at this link](https://www.kaggle.com/code/zfturbo/map-50-for-instance-segmentation).

## Submission File
For each test image (135 in total), you need to predict a set of masks for each of the 9 pathology classes. Each test image may have multiple predicted masks. Each prediction should be written on a separate line and consist of 4 fields:

`patient_id` - image ID (filename without extension)
`class_id` - class ID (ranging from 0 to 8, inclusive)
`confidence` - model confidence score
`poly` - polygon defined by a set of points in the format: x1 y1 x2 y2 ... xN yN. Each point is normalized to the [0, 1] range. Polygons follow the YOLO format specification.

The submission file must adhere to the following format:

```csv
patient_id,class_id,confidence,poly
test_000,0,0.963867,"0.100101 0.900101 0.100102 0.101010 0.901011 0.101012 0.901013 0.901041"
test_001,1,0.845342,"0.200101 0.800101 0.200102 0.201010 0.801011 0.201012 0.801013 0.801041"
...
```