Coco2Yolo Conversion[Bug]

[dexception]

I am using the following function to convert coco dataset bbox to yolo format. So for image 000000002639.jpg => https://ibb.co/b99mFd In coco dataset the annotation given is width: 640x376 5 buses with the bbox

0.0 173.34 6.34 60.75 268.69 130.97 92.1 151.24
362.54 26.16 252.35 317.96
5.1 168.19 53.25 90.62
55.89 24.98 218.49 332.81

        def convert(size, box):
            dw = 1./size[0]
            dh = 1./size[1]
            x = (box[0] + box[1])/2.0
            y = (box[2] + box[3])/2.0
            w = box[1] - box[0]
            h = box[3] - box[2]
            x = x*dw
            w = w*dw
            y = y*dh
            h = h*dh
            return (x,y,w,h)

And after conversion the coordinates generated by the convert method are: 0.004953125 0.31128989361702125 0.00990625 -0.2994414893617021 0.2818671875 0.37527925531914896 -0.27592187500000004 0.053909574468085135 0.48038281250000003 0.4576063829787234 -0.17217187500000006 0.7760638297872339 0.04558593750000001 0.3441622340425532 0.075234375 -0.2063031914893617 0.21435937500000002 0.47578457446808514 0.2540625 0.8186968085106382

But when i check in Yolo_Mark these values are not correct ! https://ibb.co/m8Qq8y

[AlexeyAB]

0.004953125 0.31128989361702125 0.00990625 -0.2994414893617021

Height can't be less than zero.

What Python script do you use?

[dexception]

The file in scripts folder voc_label.py contains the convert method but the correct method for coco dataset should be:

def convert(size, box):
    box[2] = box[2]+box[0]
    box[3] = box[3]+box[1]
    dw = 1./size[0]
    dh = 1./size[1]
    x = (box[0] + box[1])/2.0
    y = (box[2] + box[3])/2.0
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x*dw
    w = w*dw
    y = y*dh
    h = h*dh
    return (x,y,w,h)

[AlexeyAB]

http://cocodataset.org/#format-data

1. Object Detection ...

annotation{ "id" : int, "image_id" : int, "category_id" : int, "segmentation" : RLE or [polygon], "area" : float, "bbox" : [x,y,width,height], "iscrowd" : 0 or 1, }

Convertion from MS COCO-labels to Yolo-labels:

yolo_x = box[0] / image_width
yolo_y = box[1] / image_height
yolo_width = box[2] / image_width
yolo_height = box[3] / image_height

---

Why do you want to use voc_label.py to get labels of COCO-dataset? voc_label.py is of PascalVOC-dataset only.

Why do not you want to use this script to get COCO labels? https://github.com/AlexeyAB/darknet/blob/master/scripts/get_coco_dataset.sh

As described here: https://pjreddie.com/darknet/yolo/

Get The COCO Data To train YOLO you will need all of the COCO data and labels. The script scripts/get_coco_dataset.sh will do this for you. Figure out where you want to put the COCO data and download it, for example:

cp scripts/get_coco_dataset.sh data
cd data
bash get_coco_dataset.sh

[dexception]

Isn't this old ? There have been many corrections in the coco dataset.

[AlexeyAB]

Yes, if you want to use the newest COCO-dataset version, then use it from: http://cocodataset.org/#download

Convertion: http://cocodataset.org/#format-data

yolo_x = box[0] / image_width
yolo_y = box[1] / image_height
yolo_width = box[2] / image_width
yolo_height = box[3] / image_height