My json file doesn't seem to be a legit COCO dataset

I have a dataset consisting of 2164 pairs of greyscaled+groundtruth (2 instances) images. The idea would be to use detectron2 for instance segmentation, for this the masks have to be convert into COCO dataset format. Following the detectron2 specs, a json file for one image and two instances was generated as a toy model using compressed rle encoding provided by pycocotools.
Unfortunately, the dataset failed to be loaded with pycocotools or with detectron2.
The conversion was performed with a notebook available on kaggle.
Is there something wrong with the way the json dataset was generated?
For the beginning, the task would only be capable to overlay a legit COCO annotation on the greyscaled image.
Thank you for your help.