Computer Vision News 42 The success of the knee arthroplasty procedures isn’t solely reliant on the technology but also the surgeon’s expertise and experience. Pixee’s innovation complements this by empowering surgeons to implement their skills more precisely, but does it improve the success rate? “Something funny we say is that our system allows a bad surgeon to be precisely bad!” Romain laughs. “That’s not a correct answer to your question, but success rate depends on the surgeon’s knowledge, skills, and experience. Our system allows them to achieve what they’ve been planning.” Pixee’s localization algorithm is based on the ArUco fiducial marker library, and the marker pattern is designed to ensure submillimetric precision at a range of 40 centimeters, with a 4K RGB camera embedded in off-the-shelf smart glasses. Whilst the specifics of the algorithm remain under wraps, Romain shares that the pattern contains more anchor points than traditional markers. To his knowledge, this is the first time anyone has performed these kinds of in-house improvements on the ArUco library. “At some point, we hit a ceiling in terms of the precision we could achieve,” he recalls. “It pushed us to improve the ArUco concepts. We’re based on ArUco technology but slightly modified the marker pattern to get more anchor points, allowing us to reach the precision we need for knee surgery and other joint replacements.” One of the biggest technological challenges was defining a reproducible model for the lowcost, low-quality sensor embedded in the smart glasses. It is not professional or industrial-grade technology, and the team struggled to get it to produce reliable results. Each pair of smart glasses has slightly different sensor parameters that must be recalibrated over time. “It’s a lot of trial and error,” Romain reveals. “This is mostly a lot of mathematical and computer vision tricks to get better contrast and edges on the pattern we want to detect. There are a lot of layers in our image treatment to get very precise corner positions from our markers. What makes our technology unique is the combination of every little trick we have to do to achieve this precision, but mostly, it’s a tremendous amount of trial and error and testing.” AI in the Real World
RkJQdWJsaXNoZXIy NTc3NzU=