One of the key benefits of Synthetic Data is in its ability to provide both fully accurate, and advanced annotations that are not possible or very difficult and expensive to achieve with real data.
Many applications benefit from a precise understanding of a scenario, not limited to bounding boxes, but a full instance semantic segmentation, where the precise pixel masks for every desired item are marked. Chameleon provides this full synthetic data, in a pixel-perfect manner. For real world annotation by humans, re-identifying the same object/vehicle/human at a later period is near impossible, for Chameleon, it will automatically be re-identified.
The ability to provide surface normal data allows for sophisticated post-processing of generated images. The ability to understand this allows for accurate modeling of real world effects that would not be possible with "real" data
The addition of advanced annotations permits systems designers far more flexibility when training networks, especially those which consider inputs from multi-sensors such as ToF, Lidar, as well as the more usual visual sensors. Only Synthetic data can provide these advanced annotations, accurately and at a low cost
