Coco-like-dataset-creator

Convert Data created with labelme to coco format data set. This repo is aimed at creating a full dataset which include the following subsets:

    1. train set ~60% of the input images by default  
    2. test set ~20% of the input images bt default 
    3. validation set ~20% of the input images.

The output has been tasted (so far all I can tell is that the training began) on facebookresearch/maskrcnn-benchmark for instance segmentation.

create the class with the file: CreateDastaSet.py

Installation

I have based this script on the labelme labelme2coco script therefore follow the labelme installation instraction at wkentaro/labelme

How to Create the Dataset

Use the wkentaro/labelme to annotate your data
Edit CreateDataSet inputDict as follows -input = path\to\your\Inputfolder at the following format : (Bug therefore the fornat is required )
- Inputfoder -subfolder1 - containing data - Imgs and annotations
  -subfolder2 - containing data - Imgs and annotations -subfolder3 - containing data - Imgs and annotations
- outputfoler= path\to\where\output\data\will\be\created
- labels = 'path\to\label.txt'
1. Verify the output with the following notebook from [pycococreator] (https://github.com/waspinator/pycococreator)
run the CreateDataSet.py script

TODO:

1. add annotations field BoundingBox
2. convert to work with iscrowded =0 , aka polygons dots
3. add support to appending exsiting dataset (MUST Backup Previous version to dir\old\setname+timeframe )
4. re-enble parser
5. solve directory in directory bug for list of json files gathering
6. add a change directory function to \usedforcreatingdataset of input images in order to avoid using them for appending the dataset.
7. use imantics - Annotation to calculate BBox and area in stade of labelme function
8. figure how to add rotated boundingboxes

Knowon Bugs :

solved - appended empty annotations to the sets
Input dircteroy has to be parent directory bug
Image Resize doesn't mach with mask

credits:

labelme

pycococreator