🎉 Supervisely November release: reimagine image annotation and data science workflows
We at Supervisely are committed to continually improving the experience for users in our never-ending quest for perfection. Thanks to valuable feedback from customers and community, today we are excited to announce a massive upgrade: many new features, important enhancements and completely new user interfaces that bring increased usability and efficiency to your deep learning journey.
Below are a few glimpses of what’s new in this release.
New level of collaboration: teams, members, roles
A major new improvement is the ability to organize users into teams, making it easy to collaborate across multiple projects. Intuitive roles and permissions let team owners and administrators manage member access to the team’s data, computational resources and models. Now the process of configuring access to data is super transparent. With new workspaces it is now straightforward to organize and navigate between a ton of experiments.
We have also added a much-requested ability to share data and models using sharable links just like in Google Drive.
Annotation at scale: labeling jobs and activity monitoring
Large and high-quality training datasets are key to accurate neural networks. But it is challenging to organize a robust image annotation process when many workers are involved. We work with enterprises that deal with data labeling every day. In close cooperation with leading image annotation services we have accumulated best practices and developed a powerful but simple paradigm to manage image annotation effectively.
Just invite new members with role “Annotator” to your team and define Annotation Manual in project’s readme (markdown format). Now you can send necessary images for annotation as Labeling Jobs and distribute them across your labeling team. “Annotators” have access only to images that were assigned to their particular Labeling Job. So all your data and models are private.
Management of entire process is transparent. Detailed user activity and statistics allow you to see who is doing what in real time.
Best in class annotation tools for any computer vision application
Supervisely is specially designed to cover the most popular computer vision tasks such as segmentation, detection, classification, object tracking, OCR, key points and more. Our annotation platform combines best industry practices with many years of experience. Valuable feedback from community and our customers allows us to improve user experience and make work more productive over the time.
Train and evaluate custom Neural Networks on your data
Inside Supervisely you can train, evaluate and deploy state of the art neural networks in a few clicks right on your data. We have a Model Zoo with many pre-trained models. Use them with transfer learning to quickly bootstrap your custom model training. Keep your research structured, track and reproduce every experiment.
Our Model Zoo covers most popular computer tasks such as semantic segmentation, object detection, classification, text detection and OCR. Here is the list of available models (link to cookbook chapter in docs):
- YOLO v3 (DarkNet)
- DeepLab v3+ (Tensorflow)
- Faster R-CNN with different backbones (Tensorflow)
- SSD with different backbones (e.g. SSD MobileNet) (Tensorflow)
- Mask R-CNN (Keras)
- UNet v2 (PyTorch)
- ResNet, EAST, CTPN, CNN-LSTM-CTC and others
Continuously improve quality of your models: active learning and human in the loop
There is no machine learning model that works with 100% accuracy. Every model that is used in production sometimes makes mistakes. With Supervisely we can go through label-train-evaluate loop within the single environment quite fast just with mouse clicks: visualize NN predictions, find the edge cases and manually correct them. In this way you can iterate until you get both: a large training dataset and an accurate neural network. Say goodbye to manual work, switching between separate tools and converting data to different formats back and forth.
Combine Deep Learning models into custom pipelines to solve your task
Many business applications require several neural networks to be combined with custom post processing logic. Supervisely infrastructure handles such use cases out of the box: you can have several absolutely different neural networks with custom data processing steps and combine them into a single pipeline with zero lines of code. You save months of engineering time and focus on your task.
Introducing Plugins: extend functionality and customize our platform for your needs
Supervisely is a developer friendly platform. With Plugins you can extend functionality with custom add-ons: integrate your own neural networks, add custom import/export formats (such as videos or DICOM files), different metrics and so on. You can do almost anything with your data and models. Just add Supervisely SDK to your script with a few lines, implement your logic and our infrastructure will automatically handle the rest for you.
Get started with Supervisely: unify data labeling, deep learning and business
Supervisely is the leading platform for entire computer vision lifecycle. It provides rich infrastructure to annotate and manage training data, build custom neural networks, continuously improve them and combine into pipelines all on a single unified platform. This completely new paradigm significantly reduces time to market for companies and makes the work of experts more productive by automating routine steps.
Save months or even years on building custom infrastructure. Try Supervisely Community Edition for free or speak with us about an Enterprise solution for your business.
