Is there a way to save a Dataset natively, without polygon conversions?

[thomasf1]

Search before asking

• [X] I have searched the Supervision issues and found no similar feature requests.

Question

Is there a way to save a Dataset natively, without polygon conversions? Would be great If there would be a way to save the dataset with the masks as a zip, all save formats (coco, yolo, voc) do some processing. Ideally there would be a way to natively save a dataset (maybe as a ZIP file).

My usecase is experimenting to get the polygon conversion when saving as yolo after a run just right and writing some custom code for it, so it would be helpful to have the "raw" state with the masks saved somewhere.

Additional

I might just have overlooked some obvious way of saving the dataset... Otherwise it is a feature request I guess.

[github-actions[bot]]

Hello there, thank you for opening an Issue ! 🙏🏻 The team was notified and they will get back to you asap.

[SkalskiP]

Hi, @thomasf1 👋🏻 Sounds interesting. Have you perhaps already thought about the structure of such a dataset? Would each mask be stored as a separate black-and-white photo? How would you store information about the class to which an object with a given mask belongs?

[thomasf1]

Well, there is already a standard/solution for coco that´s somehow not being used by roboflow: RLE (_frString)

Also, not supporting masks causes other issues: When converting to Polygons, cutouts in masks seem often to be converted as a separate polygon.

Example: A person having this arms on his hips. The triangle shape inside the arms isn´t part of the person, but surrounded by it. The way supervisions translates that into polygons is as follows:

• The outer shape of the person as a person polygon (reasonable, expected behaviour)
• The inner cutout of background as another person polygon (blatantly wrong and counterproductive)

(not sure if that qualifies as a BUG or is somehow intended behaviour?)

[thomasf1]

@SkalskiP There is some code here that should get you far: https://github.com/waspinator/pycococreator/blob/master/pycococreatortools/pycococreatortools.py

Also, in binary_mask_to_polygon, they seem to add some padding to avoid a problem supervision (and the roboflow app smart polygon) has: Creating polygons that reach to the image edges. Unfortunately, most polygons leave a 1+ px gap from the edge of the image when the masked subject goes beyond the edge of the image.

[thomasf1]

In an Ideal world, both roboflow.com and supervision would use both masks and polygons and decide intelligently which one to use.

For instance segmentation: Masks for:

• Objects with cutouts
• Objects that are overlapped by something else, cutting the mask into separate areas that are not connected (which in polygons results in several polygons, loosing the information that they are part of one Object)
• Maybe very small Objects

Polygons for everything else (maybe configureable)

[thomasf1]

@SkalskiP What do you think?

[SkalskiP]

@thomasf1 agree! Expanding COCO annotations format support is the easiest way to unlock that capability.

[thomasf1]

Would you need any help with it?

[MihaiDavid05]

Well, there is already a standard/solution for coco that´s somehow not being used by roboflow: RLE (_frString)

Also, not supporting masks causes other issues: When converting to Polygons, cutouts in masks seem often to be converted as a separate polygon.

Example: A person having this arms on his hips. The triangle shape inside the arms isn´t part of the person, but surrounded by it. The way supervisions translates that into polygons is as follows:

• The outer shape of the person as a person polygon (reasonable, expected behaviour)
• The inner cutout of background as another person polygon (blatantly wrong and counterproductive)

(not sure if that qualifies as a BUG or is somehow intended behaviour?)

In an Ideal world, both roboflow.com and supervision would use both masks and polygons and decide intelligently which one to use.

For instance segmentation: Masks for:

• Objects with cutouts
• Objects that are overlapped by something else, cutting the mask into separate areas that are not connected (which in polygons results in several polygons, loosing the information that they are part of one Object)
• Maybe very small Objects

Polygons for everything else (maybe configureable)

@thomasf1 I opened an issue specifically for this hole preservation matter. I also built an exporter that deals with almost everything that you said above, You can find it here.

[thomasf1]

@MihaiDavid05 Great :)

One thing I could not quite work out from the ReadMe:

• From where to what does it export? I assume from Masks in Image/PNG format to Coco Masks(RLE annotations), right?

So, the way it works is Mask image (one per class? Does it suppoort multiple classes? Or Inststance Segmentation?) -> Coco Masks (RLE) if the object has holes or multiple regions, otherwise Polygons

[MihaiDavid05]

@thomasf1 Yes, that's true, from Image/PNG format to Coco Masks (rle or polygons) annotations. It supports instance segmentation, therefore multiple classes and multiple instances of the same class in an image. I will update the readme!

[thomasf1]

@MihaiDavid05 Out of curiosity, what tool did you use to generate the masks? We´re currently using hasty.ai which only allows a png masks for "Semantic Segmentation (png)". I guess using the image mask format for instance segmentation would be quite difficult...

[MihaiDavid05]

@thomasf1 Hi, I did not understand your question. Which generated masks are you talking about?

[thomasf1]

@MihaiDavid05 Sorry, I meant which tool do you use to annotate your data and in turn export that as mask images...

[MihaiDavid05]

@thomasf1, oh, I see. Currently, I'm only using already annotated base datasets, so I'm not using any tool to annotate raw images :) I might look into that!

[ryouchinsa]

We updated a script to convert the RLE mask with holes to the YOLO segmentation format. https://github.com/roboflow/supervision/issues/574#issuecomment-1821385339

[SkalskiP]

Hi, @ryouchinsa! 👋🏻 Does the YOLO format support masks with holes?

[ryouchinsa]

Hi, @SkalskiP,

Using the script general_json2yolo.py, you can convert the RLE mask with holes to the YOLO segmentation format.

The RLE mask is converted to a parent polygon and a child polygon using cv2.findContours(). The parent polygon points are sorted in clockwise order. The child polygon points are sorted in counterclockwise order. Detect the nearest point in the parent polygon and in the child polygon. Connect those 2 points with narrow 2 lines. So that the polygon with a hole is saved in the YOLO segmentation format.

def is_clockwise(contour):
    value = 0
    num = len(contour)
    for i, point in enumerate(contour):
        p1 = contour[i]
        if i < num - 1:
            p2 = contour[i + 1]
        else:
            p2 = contour[0]
        value += (p2[0][0] - p1[0][0]) * (p2[0][1] + p1[0][1]);
    return value < 0

def get_merge_point_idx(contour1, contour2):
    idx1 = 0
    idx2 = 0
    distance_min = -1
    for i, p1 in enumerate(contour1):
        for j, p2 in enumerate(contour2):
            distance = pow(p2[0][0] - p1[0][0], 2) + pow(p2[0][1] - p1[0][1], 2);
            if distance_min < 0:
                distance_min = distance
                idx1 = i
                idx2 = j
            elif distance < distance_min:
                distance_min = distance
                idx1 = i
                idx2 = j
    return idx1, idx2

def merge_contours(contour1, contour2, idx1, idx2):
    contour = []
    for i in list(range(0, idx1 + 1)):
        contour.append(contour1[i])
    for i in list(range(idx2, len(contour2))):
        contour.append(contour2[i])
    for i in list(range(0, idx2 + 1)):
        contour.append(contour2[i])
    for i in list(range(idx1, len(contour1))):
        contour.append(contour1[i])
    contour = np.array(contour)
    return contour

def merge_with_parent(contour_parent, contour):
    if not is_clockwise(contour_parent):
        contour_parent = contour_parent[::-1]
    if is_clockwise(contour):
        contour = contour[::-1]
    idx1, idx2 = get_merge_point_idx(contour_parent, contour)
    return merge_contours(contour_parent, contour, idx1, idx2)

def mask2polygon(image):
    contours, hierarchies = cv2.findContours(image, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_TC89_KCOS)
    contours_approx = []
    polygons = []
    for contour in contours:
        epsilon = 0.001 * cv2.arcLength(contour, True)
        contour_approx = cv2.approxPolyDP(contour, epsilon, True)
        contours_approx.append(contour_approx)

    contours_parent = []
    for i, contour in enumerate(contours_approx):
        parent_idx = hierarchies[0][i][3]
        if parent_idx < 0 and len(contour) >= 3:
            contours_parent.append(contour)
        else:
            contours_parent.append([])

    for i, contour in enumerate(contours_approx):
        parent_idx = hierarchies[0][i][3]
        if parent_idx >= 0 and len(contour) >= 3:
            contour_parent = contours_parent[parent_idx]
            if len(contour_parent) == 0:
                continue
            contours_parent[parent_idx] = merge_with_parent(contour_parent, contour)

    contours_parent_tmp = []
    for contour in contours_parent:
        if len(contour) == 0:
            continue
        contours_parent_tmp.append(contour)

    polygons = []
    for contour in contours_parent_tmp:
        polygon = contour.flatten().tolist()
        polygons.append(polygon)
    return polygons 

def rle2polygon(segmentation):
    if isinstance(segmentation["counts"], list):
        segmentation = mask.frPyObjects(segmentation, *segmentation["size"])
    m = mask.decode(segmentation) 
    m[m > 0] = 255
    polygons = mask2polygon(m)
    return polygons

The RLE mask.

The converted YOLO segmentation format.

To run the script, put the COCO JSON file coco_train.json into datasets/coco/annotations. Run the script. python general_json2yolo.py The converted YOLO txt files are saved in new_dir/labels/coco_train.

Edit use_segments and use_keypoints in the script.

if __name__ == '__main__':
    source = 'COCO'

    if source == 'COCO':
        convert_coco_json('../datasets/coco/annotations',  # directory with *.json
                          use_segments=True,
                          use_keypoints=False,
                          cls91to80=False)

To convert the COCO bbox format to YOLO bbox format.

use_segments=False,
use_keypoints=False,

To convert the COCO segmentation format to YOLO segmentation format.

use_segments=True,
use_keypoints=False,

To convert the COCO keypoints format to YOLO keypoints format.

use_segments=False,
use_keypoints=True,

This script originates from Ultralytics JSON2YOLO repository. We hope this script would help your business.

[ryouchinsa]

Thanks for reviewing our script. We checked whether YOLO can train polygon masks with holes with a small dataset.

Donut images and YOLO segmentation text files to confirm that YOLO can train polygon masks with holes.

[thomasf1]

Thanks @ryouchinsa. Having a look currently and it seems to work great :)

[SkalskiP]

Hi, @ryouchinsa 👋🏻 Thanks a lot for that! 🙏🏻 Making that doable with Supervision is definitely on our roadmap. We simply do not have enough capacity to take care of it now.

If any of you would like to help us out an contribute, I would be really grateful.

[thomasf1]

@SkalskiP @ryouchinsa I´ve tested the code and incorporated it into supervision. Working on a PR.

[thomasf1]

Added the PR (excuse the sample image): https://github.com/roboflow/supervision/pull/630

[ryouchinsa]

Hi @SkalskiP, I am sorry for late reply. I was working on the PR for ultralytics/JSON2YOLO. https://github.com/ultralytics/JSON2YOLO/pull/61 Now I started working on your supervision code to implement the COCO RLE to YOLO feature.

Hi @thomasf1, thanks for implementing my code onto supervision. I will read and check with my dataset.

[ryouchinsa]

I am trying to implement features which are implemented on JSON2YOLO. The script has merge_multi_segment() and merges multiple polygons into one.

Does supervision correspond to multiple polygons in the COCO format? It looks that it does not work and has an error in the function coco_annotations_to_detections() in the script supervision/dataset/formats/coco.py.

polygons = [
    np.reshape(
        np.asarray(image_annotation["segmentation"], dtype=np.int32), (-1, 2)
    )
    for image_annotation in image_annotations
]

COCO file with multiple polygons.

"annotations": [
    {
        "area": 594425,
        "bbox": [328, 834, 780, 2250],
        "category_id": 1,
        "id": 1,
        "image_id": 1,
        "iscrowd": 0,
        "segmentation": [
            [495, 987, 497, 984, 501, 983, 500, 978, 498, 962, 503, 937, 503, 926, 532, 877, 569, 849, 620, 834, 701, 838, 767, 860, 790, 931, 803, 963, 802, 972, 846, 970, 896, 969, 896, 977, 875, 982, 847, 984, 793, 987, 791, 1001, 783, 1009, 785, 1022, 791, 1024, 787, 1027, 795, 1041, 804, 1059, 811, 1072, 810, 1081, 800, 1089, 788, 1092, 783, 1098, 784, 1115, 780, 1120, 774, 1123, 778, 1126, 778, 1136, 775, 1140, 767, 1140, 763, 1146, 767, 1164, 754, 1181, 759, 1212, 751, 1264, 815, 1283, 839, 1303, 865, 1362, 880, 1442, 902, 1525, 930, 1602, 953, 1640, 996, 1699, 1021, 1773, 1039, 1863, 1060, 1920, 1073, 1963, 1089, 1982, 1102, 2013, 1107, 2037, 1107, 2043, 1099, 2046, 1097, 2094, 1089, 2123, 1074, 2137, 1066, 2153, 1033, 2172, 1024, 2166, 1024, 2166, 1023, 2129, 1019, 2093, 1004, 2057, 996, 2016, 1000, 1979, 903, 1814, 860, 1727, 820, 1647, 772, 1547, 695, 1637, 625, 1736, 556, 1854, 495, 1986, 459, 2110, 446, 1998, 449, 1913, 401, 1819, 362, 1720, 342, 1575, 328, 1440, 335, 1382, 348, 1330, 366, 1294, 422, 1248, 437, 1222, 450, 1190, 466, 1147, 482, 1107, 495, 1076, 506, 1019, 497, 1016],
            [878, 2293, 868, 2335, 855, 2372, 843, 2413, 838, 2445, 820, 2497, 806, 2556, 805, 2589, 809, 2622, 810, 2663, 807, 2704, 793, 2785, 772, 2866, 742, 2956, 725, 3000, 724, 3013, 740, 3024, 757, 3029, 778, 3033, 795, 3033, 812, 3032, 812, 3046, 803, 3052, 791, 3063, 771, 3069, 745, 3070, 733, 3074, 719, 3077, 702, 3075, 680, 3083, 664, 3082, 631, 3072, 601, 3061, 558, 3058, 553, 3039, 558, 3023, 566, 3001, 568, 2983, 566, 2960, 572, 2912, 571, 2859, 567, 2781, 572, 2698, 576, 2643, 583, 2613, 604, 2568, 628, 2527, 637, 2500, 636, 2468, 629, 2445, 621, 2423, 673, 2409, 726, 2388, 807, 2344, 878, 2293]
        ]
    }],

The script which converts COCO to YOLO using supervision.

import supervision as sv

sv.DetectionDataset.from_coco(
    images_directory_path= r"/Users/ryo/rcam/test_annotations/test/_test_min_polygon",
    annotations_path=r"/Users/ryo/rcam/test_annotations/test/_test_min_polygon/coco_train.json",
    force_masks=True
).as_yolo(
    images_directory_path=r"/Users/ryo/rcam/test_annotations/test/_test_min_polygon/move1",
    annotations_directory_path=r"/Users/ryo/rcam/test_annotations/test/_test_min_polygon/move2",
    data_yaml_path=r"/Users/ryo/rcam/test_annotations/test/_test_min_polygon/data.yaml"
)

[SkalskiP]

We just opened a new issue https://github.com/roboflow/supervision/issues/1114 where we proposed adding RLE support to Supervision datasets. To keep it clean and prevent duplications, I'm closing this issue. Also, if any of you would be willing to help us out with implementation, let us know! 🙏🏻

[SkalskiP]

@ryouchinsa we do not support disjointed masks for now, but we are actually working on the PR that may add this https://github.com/roboflow/supervision/pull/1086