Data annotation takes the most effort in today’s machine vision model development. Deep learning models often need millions of images for training to achieve the necessary accuracy and generalization. This indicates that a huge amount of manual work would be needed for gathering and annotating the images. We have created a solution to make this process easier and more efficient.
At Vaisto, we are applying hybrid transfer learning methods and synthetic baseline training images with auto-annotation generation to overcome the challenges of creating massive datasets. In order to create a realistic environment in the 3D world, special simulation applications are to be annotated.
Existing deep learning architectures can be used and potentially pretrained weights as well. When first training with simulated images, the model only learns to detect the target objects in the virtual environment. Training image augmentation, where images are manipulated randomly during the training, is used to narrow the gap between simulated and real-world pictures. These are needed to fine-tune the model to start detecting objects in the real world.
It creates circumstances or events that are otherwise difficult to reproduce. This is the importance for the model to become generalized. There’s a risk that real images must be annotated manually, but there’s also a big possibility that we can use the synthetic image to support semi-automatic annotation, such as in active learning.
Real-world images need annotations. Generating and annotating pictures automatically saves a lot of time, effort and energy. Vaisto provides the solution to make the process faster, easier and less error prone.
Hermiankatu 6-8 L
33720 Tampere, Finland