GTC 2020: Raytracing for Deep Neural Networks Training
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Raytracing for Deep Neural Networks Training
Jakub Pietrzak, Sky Engine Limited
Learn how to prepare virtual scene and render images dedicated to train deep convolutional neural networks. Make sure that the model trained in virtual reality can be efficient and accurate in the real world. We present Sky Engine, the very first raytracer designed for data generation for deep neural networks training. Our NVIDIA OptiX based raytracer is tightly integrated with the PyTorch framework and can render images, semantic masks, and depth fields on the fly. Sky Engine can change scene parameters during the training process to balance the dataset automatically and detect most scene configurations that will confuse the model.