How to make my own dataset in COCO format for MMFewshot？

[cmjkqyl]

Referring to README, the data is organized as follows: mmfewshot ├── mmfewshot ├── tools ├── configs ├── data │ ├── coco │ │ ├── annotations │ │ ├── train2014 │ │ ├── val2014 │ │ ├── train2017 (optional) │ │ ├── val2017 (optional) │ ├── few_shot_ann │ │ ├── coco │ │ │ ├── annotations │ │ │ │ ├── train.json │ │ │ │ ├── val.json │ │ │ ├── attention_rpn_10shot (for coco17) │ │ │ ├── benchmark_10shot │ │ │ ├── benchmark_30shot

For the folder new_shot_ann, I'm not sure what kind of content should be placed. My understanding of this data organization is that data/coco holds the complete dataset with all categories. few_shot_ann/coco/annotations holds the annotations for the full support set and query set. benchmark_10shot and benchmark_30shot hold the annotations for the query set. Is this a correct understanding? I want to extract the contents of train.json and val.json in new_shot_ann from instances_train2017.py and instances_val2017.py in my own industrial dataset, how should I do it?