PytorchCV Studio

Inspiration

Computer vision is increasingly becoming part of our day-to-day lives. Doctors use deep learning models to accurately identify acute illnesses, self-driving cars use them to avoid obstacles, and your little sister uses them to apply Instagram filters. While these applications are life-saving, impressive, and fun, they only became so with a human teacher. Tools for annotating images for computer vision exist, but highly flexible tools typically come at a price that many independent developers can’t afford. We seek to bring the full potential of computer vision to the greater community by providing a set of tools to address the most common tasks and pairing them with Pytorch’s deep learning resources.

What it does

PytorchCV allows users to easily annotate images using a label system for classification problems, a bounding box for localization, and a polygon tool or freehand selection for segmentation tasks. Once your annotations have been made, PytorchCV uses Pytorch’s resources to train a deep learning model. To do so, the user opens the experiment station. Here, they select their training dataset and specifies the type of problem they want their model to solve. To take the community interaction to the next level, PytorchCV can also export your trained model to the Pytorch Hub. Here, other users can browse, use, and continue to improve computer vision models trained by other users. By providing the tools for the open-source community, we hope to become the standard for those working on computer vision problems.

Challenges we ran into

The idea behind the software is offering a tool for the PyTorch developers and the community in general, to prepare the datasets, train, evaluate and make the models available to the world taking advantage of the PyTorch hub platform using a friendly user interface.