Computer Vision News 40 RSIP Vision’s Dekel Shapira and Artium Dashuta talk to us about their work with Blender, a free and open-source 3D creation suite, to generate synthetic video data that can be used for various Machine Learning projects. Dekel and Artium chose the 3D content creation software Blender when a recent project required the creation of realistic synthetic videos of surgical procedures. Other programs can render 3D objects and scenes, but the team needed a tool that could seamlessly integrate 3D assets to generate videos that look like real cases. By knowing the clinical use-case they generated a basic scene. This scene was passed to a medical expert annotator, who moved the tools authentically before it was rendered in Blender. “Whenever you train a neural network, you want some ground truth data,” Dekel tells us. “You need pictures that are as close as possible to what you see in real cases. We saw many videos of real procedures, identified the stages and created a guide to what should appear in the blender scenes”. Despite its impressive features, using Blender was not without its challenges. Dekel says it is not the ideal tool for programming. “It has a Python Interface, but it’s not very convenient to debug and not very persistent in the API,” he reveals. “Also, it has quite a high learning curve. Once you know it, it’s very convenient, but it takes time to learn it because it’s very different from other programs. It has its own logic”. The creation of the 3D assets is also non trivial and sometimes can required the assistance of a 3D artist. This is worthwhile, since sometimes the rendered results can look so realistic that even an expert will find it challenging to discriminate between a synthetic image and a real one. RSIP Vision’s MedTech Projects
RkJQdWJsaXNoZXIy NTc3NzU=