Collecting data is hard
A common challenge when building and training neural networks is that they often need to be trained using large volumes of training data and collecting this data can be very hard. In this project we created a pipeline to generate large amounts of 3D rendered images to supplement the existing training data. By training on a data set consisting both of real world images and synthetic images the robustness and accuracy of a model can be improved.
Closing the reality gap
One problem with using rendered images instead of photos to train a model is the “reality gap”. Even if 3D rendering has come a long way there are still some differences between photos and rendered images that could impact the model negatively and make it perform worse on photographs. In the project we used several techniques to make the synthetic images more “realistic” and one of these techniques was by compositing 3d models into photographs and simulating lighting, shadows and noise.
Privacy is important
When working on image analysis it is important to take privacy into account, especially when collecting training data. Synthetic images don’t contain any personal information and thus makes it possible to create large data sets.