hachreak / subcoco

Build a subset of COCO dataset

Stats

If you want collect multiple statistics from a COCO-like dataset (output is a csv file):

$ python stats.py
usage: stats.py [-h] [-d DELTA] train_json val_json {instances_per_category,cats_per_img,instances_per_img,instance_size}
stats.py: error: the following arguments are required: train_json, val_json, action
python stats train.json val.json

Actions:

• instances_per_category: instances per category - count how many instances per class (also with percentage).
• cats_per_img: categories per image - number of annotated categories per image (% of images / number of categories).
• instances_per_img: instances per image - number of annotated instances per image (% of images / number of instances).
• instance_size: instance size - the distribution of instance sizes. Option: -d 0.1 specify delta percentage value.

View

Paint images with bbox and segmentation annotations:

$ python view.py annotation.json images_dir output_dir

SubCoco

Get a subset of COCO.

Example: show categories stats on screen:

python subcoco.py $COCODATASET/annotations/instances_val2017.json \
        -l data/coco.names --stats

Example: you want only categories motorcycle, backpack, sheep and cow in your subset of coco dataset.

python subcoco.py $COCODATASET/annotations/instances_train2017.json \
        -c 4 20 21 27 -o $COCODATASET/annotations/instances_subtrain2017.json

Example: print category ids.

python subcoco.py $COCODATASET/annotations/instances_train2017.json -i

Help:

python subcoco.py -h