merge_from_file.md

Merge From File

This is a way of merging datasets (each in coco format) based on a list file.

merge_list is a txt file which contains a list of coco dataset directories to merge. Each coco dataset directory should have an images folder and a <directory name>.json annotation file. In the source list file, dataset split modes are indicated by [square brackets], followed by the constituent directory names. An empty lines or new square bracket will indicate the end of the previous split. Directory names must be unique.

For e.g.:

[train]
iphone_dataset1
iphone_dataset2
dslr_dataset1
# comments
dslr_dataset2

[val]
iphone_dataset3
dslr_dataset3

Comments ("#" preceding line) are ignored.

root is the parent folder containing all the coco dataset folders (can be nested, but dataset folder names have to be unique). If not given, will search within merge_list's folder.

This will merge all categories in all given datasets.

Merging logic follows the usual merge.

Usage

python3 -m cocojson.run.merge_from_file -h

usage: merge_from_file.py [-h] [--root ROOT] mergelist output

positional arguments:
  mergelist    Path to merge list
  output       Path to output directory

optional arguments:
  -h, --help   show this help message and exit
  --root ROOT  Path to parent folder containing all the coco dataset folders
               to be merged (optional, will search mergelist directory if not
               given.)

Examples

python3 -m cocojson.run.merge_from_file /data/datasets_split.txt  /data/test_collated/