Cocotools

introduce

A python lib with shell entry for managing json dataset files in coco format, support detection and keypoint working

install

• python3 -m pip install cocotools
• coco --install-completion (using this for better shell completion)

uninstall

• python3 -m pip uninstall cocotools

usage

create coco file

1. create empty coco file from image directory

    coco from-image-dir IMAGE_DIR [--with-box] [-o OUTPUT_FILE]

2. create coco file from human labeling result

    coco convert-box-labeling LABEL_FILE IMAGE_DIR [-o OUTPUT_FILE]

   The labeling file should be like this:

   {
       "img_name_1.jpg": {
           "data": [
               {
                   "bbox": [
                       2.598000000000013,
                       97.862,
                       152.422,
                       155.886
                   ],
                   "type": "car",
                   "values": {},
                   "id": 1
               },
               {
                   "bbox": [
                       176.67099999999996,
                       114.31700000000001,
                       160.217,
                       129.905
                   ],
                   "type": "car",
                   "values": {},
                   "id": 2
               }
           ]
       },
       "img_name_2.jpg": {
           "data": [
               {
                   "bbox": [
                       0,
                       508.394,
                       335.324,
                       497.577
                   ],
                   "type": "person",
                   "values": {},
                   "id": 1
               }
           ]
       }
   }

visualize

Visualize box (and keypoints) result on detection (or ground truth) files

coco visualize COCO_FILE IMG_DIR

print stats

Show how many images/boxes/categories in a coco file

coco print-stat COCO_FILES...

evaluation

coco evaluate [OPTIONS] GT_FILE DT_FILE

merge different coco files

coco merge [-o, --output FILE] INPUTS_COCOS...

convert id type

1. using string id

   coco to-str-id COCO_FILE

2. using integer id

   coco to-num-id COCO_FILE

split dataset

split dataset into train (80%) and val (20%)

coco split-dataset COCO_FILE IMAGE_DIR

others

1. coco cmd "SOME PYTHON CODE" using python code for more operations
2. python -m coco.badcase GT_COCO DT_COCO -i IMAGE_DIR for visualizing badcase //1. python -m coco.distribution_analyze ref coco for obtain channels which boxes similar with ref's
3. from coco import COCO using lib in python code

changelog

• 0.2.0.1 using pypi for distribution
• 0.2.0.0 add badcase module for badcase box visualize add distribution_analyze module for analyze box distribution in one dataset
• 0.1.1.5 add keep_images option for filters
• 0.1.1.4 add debug argument
• 0.1.1.3 support detection file with num ids
• 0.1.1.2 add -e for evaluate
• 0.1.1.1 fix console entry
• 0.1.1 add argparse
• 0.1.0 initialize the repo