[BUG] COCO export: Instance Segmentations OOB wrt their bounding box

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 20.04
• Python version (python --version): 3.10.10
• FiftyOne version (fiftyone --version): 0.20.0
• FiftyOne installed from (pip or source): pip

Commands to reproduce

def check_ann(ann):
    bb = ann['bbox']
    segall = ann['segmentation']

    bbxyxy = bb.copy()
    bbxyxy[2] += bbxyxy[0]
    bbxyxy[3] += bbxyxy[1]

    for seg in segall:
        seg = np.array(seg,dtype=np.float64).reshape((len(seg)//2),2)

        #print(seg1)
        orx0 = seg[:,0]<bbxyxy[0]
        orx1 = seg[:,0]>bbxyxy[2]
        ory0 = seg[:,1]<bbxyxy[1]
        ory1 = seg[:,1]>bbxyxy[3]
        if np.any(orx0):
            print(f'x0, {seg[orx0,0]} against {bbxyxy[0]}')
        if np.any(orx1):
            print(f'x1, {seg[orx1,0]} against {bbxyxy[2]}')
        if np.any(ory0):
            print(f'y0, {seg[ory0,1]} against {bbxyxy[1]}')
        if np.any(ory1):
            print(f'y1, {seg[ory1,1]} against {bbxyxy[3]}')
    
    return seg, bbxyxy

PATH = '/home/allen/tmp/export_oiv7'

ds = foz.load_zoo_dataset('open-images-v7',max_samples=15,
                          split='validation',label_types=['segmentations'])

ds.export(export_dir=PATH,dataset_type=fo.types.COCODetectionDataset)

ds2 = fo.Dataset.from_dir(dataset_dir=PATH,dataset_type=fo.types.COCODetectionDataset)

JSON = os.path.join(PATH,'labels.json')
with open(JSON) as fh:
    j0 = json.load(fh)

for aidx,ann in enumerate(j0['annotations']):
    print(aidx)
    check_ann(ann)

Output:

0
x0, [752.5] against 753.651712
1
2
3
y0, [154.5 153.5] against 154.77248
4
5
y0, [350.5 351. ] against 351.99974399999996
6
x0, [266.5 266.5] against 267.327488
y0, [166.5 167.5] against 167.727725
7
y0, [86.5] against 87.560681
8
y0, [71.5 71.5] against 72.42499
9
y0, [468.5] against 469.25214600000004
10
x0, [102.5 103.5] against 103.999488
y0, [86.5 86.5] against 87.999744

... snip ...

Describe the problem

Exporting instance segmentations via COCODetectionDataset can result in segmentations that go out of bounds with respect to their bounding box by up to ~1.5px. This may cause an issue for some training pipelines and/or represent a small loss in accuracy.

I investigated one of the original COCO datasets downloaded from cocodataset.org and this issue was not present. So round-tripping through FO (importing and then re-exporting) likely causes the slight discrepancies.

What areas of FiftyOne does this bug affect?

• App: FiftyOne application issue
• Core: Core Python library issue
• Server: FiftyOne server issue

Willingness to contribute

The FiftyOne Community encourages bug fix contributions. Would you or another member of your organization be willing to contribute a fix for this bug to the FiftyOne codebase?

• Yes. I can contribute a fix for this bug independently
• Yes. I would be willing to contribute a fix for this bug with guidance from the FiftyOne community
• No. I cannot contribute a bug fix at this time