Generation of high quality training sets for Machine Learning algorithms is recognized to be one of the most complex and labor intensive tasks in the network training process. In many cases images or video must be painstakingly hand-annotated by a human to ensure training set data of sufficient quality and fidelity for optimum neural network performance.
My Sky Technologies has augmented the expensive and labor intensive practice of developing annotated data sets with a proprietary Visual Content Creation Toolchain (VCCT). This offers a depth of capability for synthetic training data generation. This system offers:
Fully synthetic image sets
Images of highly detailed 3D models are rendered in a highly realistic virtual environment. This provides the utmost in flexibility, as practically all environmental conditions and model orientations may be manipulated automatically in software. This includes realistic season and time of day environmental data for physical locations, weather, and object orientation.
Semi-Synthetic Image sets
Similar to a fully synthetic dataset, but using highly detailed 360-degree camera images of the desired environment. 3D models are positioned within this environment in a realistic way to provide an ensemble of training images that are in the client's specific context of interest.
Using high resolution photographs, 360-degree image capture, and sophisticated computer image manipulation, virtual scenes are composited using real objects in real environments. This approach trades the flexibility of the synthetic generation methods for high detail fidelity and contextual awareness.