Wolution Blog

We have recently moved to our new office in Munich. Our new address is 

Lochhamer Str. 31 
82152 Martinsried

Martinsried is a hot spot of Biotechnology and is the home for institutions like Campus Martinsried, Ludwig-Maximilian University (LMU) München, the Max Planck Institute of Biochemistry, the Max Planck Institute of Neurobiology as well as the Innovation Center for Biotechnology (IZB). Those facilities and numerous spin-offs make Martinsried to one of the largest european centers where science, innovation and innovative businesses are merged together. We could hardly have found a better location for Wolution and look forward to local collaborations.      

A common task of our clients is to analyse stacks of images or videos, and to put the single images in relation to each other. For example some of our clients want to track the movements of animals in stables to study their behavior. This is now possible with Wolution.

As usual you first create a set of training data, where you mark the objects of interest in the frames of the video. The algorithm is then trained for the task. You can analyze arbitrarily long videos, and we support most standard video formats.  

We strive to make cutting edge machine learning technology for image analysis accessible to non-expert users. To achieve this, we keep the user interface as simple as possible and include only a few options which the user can choose. In this update we have further improved our user interface, with information boxes that explain these options on the spot where they can be selected. For example here is the box that helps you to select the right algorithm for your task:

As you can see we currently provide three classes of learning algorithm, some of which come in several versions. If you have more questions about this please contact us.

The key results of your image analyses are often lists of detected objects and their positions and properties. Wolution gives you this data in two convenient formats:

• CVS tables which you can plot and analyze in Excel (or many other CVS compatible programs).
• PDF documents which are easy to read, print and share.

We are now providing summary statistics for all images in an analysis of two different types: One file only includes global information about each image (how many objects, how big are all segmentation areas together etc.), and a second file gives you object by object statistics. In this way you always find the right data file for your task. Here is an example of parts of the detailed CSV table for object detection.

Most of our customers use segmentation to analyse their images. While this is a flexible technique, it can be tedious to create training data. If you are only interested in counting objects and don't care for their shape, you can now use bounding box detection. Bounding boxes are quick to draw in the training data editor, so you easily create a training sample that is big enough. In addition, bounding boxes, unlike classic semantic segmentation, can tell objects apart which are overlapping. We are also beta testing an instance segmentation algorithm, that combines the advantages of bounding box detection and semantic segmentation. Here is an example of bounding box detection for cars in satellite images:

Image by KIT AIS Data Set

Today we are making our new core segmentation algorithm available, which we have been beta testing for a while. Our algorithm uses a deep convolutional neural network that we train based on the ResNet101 model. The ResNet models have won multiple image recognition awards, including the important ImageNet and COCO challenges. Using these models for image segmentation (as opposed to image recognition) comes with new challenges, and we are using atrous convolutions among other tricks to get good results with manageable computational cost. Our new algorithm is outperforming the old one on all test tasks, and allows to solve very challenging new problems. We are looking forward to seeing it perform on the real world applications of our users.

Here is an example of our new algorithm, finding dolines in satellite images: