xt-cvdata

Description

This repo contains utilities for building and working with computer vision datasets, developed by Xtract AI.

So far, APIs for the following open-source datasets are included:

1. COCO 2017 (detection and segmentation): xt_cvdata.apis.COCO
2. Open Images V5 (detection and segmentation): xt_cvdata.apis.OpenImages
3. Visual Object Tagging Tool (VoTT) CSV output (detection): xt_cvdata.apis.VoTTCSV

More to come.

Installation

From PyPI:

pip install xt-cvdata

From source:

git clone https://github.com/XtractTech/xt-cvdata.git
pip install ./xt-cvdata

Usage

See specific help on a dataset class using help. E.g., help(xt_cvdata.apis.COCO).

Building a dataset

from xt_cvdata.apis import COCO, OpenImages

# Build an object populated with the COCO image list, categories, and annotations
coco = COCO('/nasty/data/common/COCO_2017')
print(coco)
print(coco.class_distribution)

# Same for Open Images
oi = OpenImages('/nasty/data/common/open_images_v5')
print(oi)
print(coco.class_distribution)

# Get just the person classes
coco.subset(['person'])
oi.subset(['Person']).rename({'Person': 'person'})

# Merge and build
merged = coco.merge(oi)
merged.build('./data/new_dataset_dir')

This package follows pytorch chaining rules, meaning that methods operating on an object modify it in-place, but also return the modified object. The exception is the merge() method which does not modify in-place and returns a new merged object. Hence, the above operations can also be completed using:

from xt_cvdata.apis import COCO, OpenImages

merged = (
    COCO('/nasty/data/common/COCO_2017')
        .subset(['person'])
        .merge(
            OpenImages('/nasty/data/common/COCO_2017')
                .subset(['Person'])
                .rename({'Person': 'person'})
        )
)
merged.build('./data/new_dataset_dir')

In practice, somewhere between the two approaches will probably be most readable.

The current set of dataset operations are:

• analyze: recalculate dataset statistics (e.g., class distributions, train/val split)
• verify_schema: check if class attributes follow required schema
• subset: remove all but a subset of classes from the dataset
• rename: rename/combine dataset classes
• sample: sample a specified number of images from the train and validation sets
• split: define the proportion of data in the validation set
• merge: merge two datasets together, returning merged dataset
• build: create the currently defined dataset using either symlinks or by copying images

Implementing a new dataset type

New dataset types should inherit from the base xt_cvdata.Builder class. See the Builder, COCO and OpenImages classes as a guide. Specifically, the class initializer should define info, licenses, categories, annotations, and images attributes such that self.verify_schema() runs without error. This ensures that all of the methods defined in the Builder class will operate correctly on the inheriting class.

Data Sources

[descriptions and links to data]

Dependencies/Licensing

[list of dependencies and their licenses, including data]

References

[list of references]