More data to establish new technologies
Due to the lack of affordable, high quality data for custom object detection and segmentation tasks there are a lot of fields being unexplored.
Our solution helps to generate meaningful synthetic images for training of neural networks in a short amount of time containing pixel wise annotation with a high, stable quality of annotations throughout the whole dataset.
Solving Computer Vision tasks with Deep Neural Networks has become state of the art in many areas in recent years. Not relying on rule based methods makes Deep Learning models more flexible and applicable to very domain specific tasks. But those models are hungry for loads of data, in our case images, which always need to be annotated by humans. Current workarounds in the field of Deep Learning such as Transfer Learning helps but still a lot of annotated data is required to develop a robust model for object detection or segmentation.
After taking images of the objects of interest those images need to be checked for quality and then sent to either a domain expert who has enough knowledge to annotate the images with the right class or to an annotating worker inside or outside the company developing the solution. For some tasks domain experts are the only source of high quality annotated ground truth data, but since their time is limited the amount of data is the bottleneck.
In other cases domain knowledge is not required, but since the annotation is done by humans the dataset is prone to human error.
Each of the solutions involves either a lot of overhead due to management of the labeling process or involves a lot of resources locked in if you take care of the whole labeling process yourself.
By combining our knowledge in the field of Deep Learning and Computer Generated Imagery we have developed a solution which is capable of creating lots of pixel wise annotated images within a short time frame.
Considering the fact that the dataset is created synthetically we assure that the dataset will be unbiased due to quality control steps during the creation process. Since a strong variation in the environment allows the neural network to focus on the important elements of the image, we assure that each image is unique.
Since trained models should work in various environments as e.g. in SaaS solutions for consumers or end customers it is nearly impossible to create a sufficient dataset which represents the variety of our real world. The bigger the project and the higher the number of objects of interests the more time it takes to first plan the creation of the dataset and more resources are locked into managing quality and working on preprocessing methods due to the lack of labels.
Synthetic images offer the best trade off between quality and price for training object detection and segmentation models.
Since our images are created by computers we not only get perfect labels on each object in the image but also have pixel wise annotations and depth maps which assures that you get the best labels possible for object detection and image segmentation tasks.
Best results can be achieved by combining a small amount of annotated real world images from the domain with a huge portion of synthetic images. The model will have the ability to generalize due to randomness in the synthetic images and memorize to some degree the domain specific feature from the real world images.