Case Study

🎉 Open Beta Release 🎉

4min read

Concept Creation

In 2017 at Munich’s Oktoberfest, the founders of Moonvision undertook the challenge of tracking dishes at one of the festival’s largest tents. To do this, we created a demo python server with a small jquery frontend to serve as our first proof of concept. This small application was able to run in realtime as dishes left the kitchen and achieve 98.2% accuracy.

We learned 2 main things from this:

  1. Tools for labelling and maintaining image and video datasets is severely lacking
  2. There is a need for computer vision technology where little to no data currently exists

The MoonVision platform was created to solve these problems. Our first step is the Open Beta Release of our annotation tool. For no cost, the annotation tool lets you and your team grow your dataset while actively learning your dataset. When you are finished labelling, you can download your dataset and train a model as you see fit or take advantage of Moonvision’s data efficient training that gets you better results with fewer labels and let us train for you.

The platform became a full service computer vision solution that takes care of the hard parts of extracting images from video, labelling images, training a model, and running that model at scale. We look forward to hearing your feedback and hope the labelling tool is as useful for you as it is for us internally.

Infrastructure

We have come a long way since our early prototyping days and now run on Google Kubernetes Engine. We use a Django server providing our main computer vision capabilities and Dask for offline processing.

With that setup, we have already processed hundreds of hours of video, labeled tens of thousands of images and automatically labelled tens of thousands of images in real time (under a half a second) using cloud GPUs. We look forward to supporting you and your data on our tested infrastructure.

An example of processing video with 85 Cores simultaniosly still gives us the chills.

User Experience

Computer Vision Systems always serve a very specific target audience. Our competition ranges from very general systems that can do everything in AI to very specific approaches in Life Science and and Medical Imaging.

Our current users are domain experts in their industry who need to automate their processes that usually needs to be done by humans. Hence, we created tools that enable such experts to create models that inspect those processes automatically.

The platform is ready to handle large datasets in terms of raw input (a camera that records video), whilst the actual amount of training data is often not that massive. When clients ask us to train a model, we use our data efficient training technology. Based on the first results, targeted labelling can constantly improve the quality and scope of the detection.

August 2017 - First Frontend for Labelling (Customized VIA Tool)

Free Open Beta Features

The current release of our open beta features focuses on data management, team management and annotation tasks. It provides a very powerful annotation tool for efficient and collaborative work with image data.

Quick Annotation Workflow

Our interface is designed to optimize labelling speed. The annotation tool is developed for power user settings with maximal usage of screen resolution, zooming, undo/redo, and hotkeys.

Powerful Recommendation Engine

The annotation process is tedious and others have come up with the solution of outsourcing the labelling process to ease this pain. This works when what you try to label doesn’t require domain expertise. We developed a recommendation engine to actively learn while you label and returns results after the first annotation given. It is our goal that teaching a computer be just as easy (if not easier) than teaching another human to label.

Collaboration

The Moonvision platform allows for multiple owners to access and label your data, but also allows you to assign labelling only permissions to a user so that they may contribute new labels to your dataset but not have access to administrative capabilities. After your users have labeled, automatic sprite generation allows you to quickly verify their results and correct any mistakes that might have been made along the way.

Premium Happiness

For our premium clients we give tools to work with video, which is still the cheapest way to generate data. We provide the process of smart scene detection, snapshot extraction with motion blur avoidance. We can generate and draw entities automatically, cluster them for bulk labelling and work with pre-trained network for better label quality.

The core of our technology stack is training with little data. Using transfer learning, our models make use of label hierarchies to provide you with the best possible output with only a small amount of training data.

Additionally, many more features are available in labelling, data management and data gathering, that gives domain experts the possibility to automate and improve the quality with little time.

Custom segmentation model training with 11 annotated images only.

Get started right away

Our goal is that the Moonvision platform is a full service computer vision platform that takes care of the hard parts of extracting images from video, labelling images, training a model, and running that model at scale. We look forward to hearing your feedback and hope the labelling tool is as useful for you as it is for us internally.

Go to app.moonvision.io/signup or talk to us for an individual consultation.