Good interview on the use of synthetic data in computer vision problems for autonomous driving.

Key points:

• Synthetics are extremely useful
• Case on how the tracker learned to work in conditions of a large number of closed objects on the simulator
• Simulators are photorealistic enough to give a boost, and in the case of heavy models without them it is almost impossible to collect the required amount of high-quality labeled data
• Very cool case of how they used multi-tasking deep learning to improve semantics where they took most of the data from the simulator and estimated the depth in real data using monocular depth network
• Idea: you can use real data only for validation, and depending on errors for some case / class, try to generate more synthetic data automatically to fix the error
• Gaidon is cool