Introducing Supervise.ly 🎉
Dataset management, annotation and preparation service
Right now Deep Learning is taking over the world. Self-driving cars, composing music, replicating an art — what seemed like magic yesterday is nothing special today. And you don’t have to be a wizard or math Ph.D to create such things. Do you have a graphic card? You’re good.
One of the main reasons why Deep Learning is not a rocket science anymore are all those frameworks. Tensorflow, Caffe, Keras — you can build complex networks as if playing with LEGO pieces.
But not a fun part of it is data, people tend to forget about it. It’s like a fuel for machine learning and without data algorithms mean nothing. There are lots of cool libraries for building neural networks, but not so many tools for data preparation.
Large businesses are able to collect large amounts of data and invest money in annotation on services like Amazon Mechanical Turk. But what small companies and researches can do?
Why did we build Supervise.ly
Here at Deep Systems we create AI solutions for various business sectors and face challenges everyday. To apply state-of-the-art models we first need to collect data, combine it with open-source datasets, annotate and prepare it for training. Usually it takes most of our time! 😫
To make things easier we had to come up with a solution. We wanted to create a place where you can go through every step of dataset preparation: from data to neural network model. And today we want to share our service called Supervisely with the world! 🎉
So how exactly Supervise.ly can help our fellow data scientists?
- Online service to work with data. Keep and access your data anywhere, with backups and handy tools.
- Datasets import. Upload your own data or automatically covert popular open-source datasets — we include many ready-to-go configurations.
- Annotation tool. Annotation of a single image takes up to 30 minutes: it’s important to have a handy and practical tool. We support hot keys, both vector and bitmap figures, filtering and captioning and a lot more in a web-based solution.
- Smart export. Keep datasets in one place and prepare them together! No more coding: use simple and plain JSON-configuration to make filtering, resizing, augmentation, train-validation splitting, combining multiple datasets in one — and save your results in popular ready to train frameworks formats.
- Statistics. Know your data well and you won’t face unexpected trouble during training. Things like class imbalance problem can make your day worse if you‘re not prepared.
- Role management. Enterprise users will appreciate multiple accounts, access permissions and activity log. For example, configure different rights for m-turkers, in-house annotators and data scientists.
- Integrations. Combine the power of well-known tools like Amazon Mechanical Turk with features of Supervisely.
- Students and researchers will like the simplicity of managing open-source datasets. Besides, for non-commercial purposes it’s absolutely free.
- Business and experts can have a single place to keep their valuable information safe and use the advantages of online annotation. If privacy is a question, no problems — you can have Supervisely at your own servers.
To make things even easier we have started a series of tutorials of how to train modern neural networks on open-source datasets like Cityscapes. The first part is already here: