Using the Power of synthetic data for advanced 3D vision
After internal developments in the creation and utilization of synthetic data, we have opened up numerous possibilities, enhancing the development of vision algorithms. Traditional methods often require extensive manual effort—taking pictures and meticulously annotating them. However, these labor-intensive processes may still fall short, producing datasets that do not accurately represent real-world scenarios.
Our 3D vision algorithms are significantly improved using synthetic data. In the video, we showcase that 3D picking of reflective objects is possible with this system. Traditional sensors struggle with accurate depth estimation of these textures, and modern systems can be prohibitively expensive. At Newcircle Technologies, we have developed a new and automated pipeline to address these challenges.
This pipeline uses 3D models, 3D object scans, or parametric models as input. Synthetic data is generated, allowing the robot to start learning in the virtual world. The deployed models are then extensively tested against various disturbances such as light shifts, background changes, rust, and other textural inconsistencies.
In summary, synthetic data is revolutionizing detection algorithms by offering a scalable, customizable, and efficient solution for developing advanced vision algorithms. Its ability to model complex scenarios and provide robust training data is driving advancements in the field of robotics.