3D Reconstruction AI Tracks the Same Person Through Scene Changes

According to AI TIMES, an UNIST research team unveiled an AI on October 15 that can reconstruct people in 3D by tracking multiple individuals as the same person even in edited videos with frequent scene cuts. The technology cut by about 62% errors that mistakenly identify the same person as someone else, and its key innovation is improving motion continuity by aligning spatial standards in video segments that show the same location. It addressed the limits of conventional methods, which were prone to shifts in camera angles and editing, by aligning the background and camera position together and then making judgments based on people’s positional relationships and appearance information. As a result, the media reported that it can produce more natural reconstructions by reducing shaking and slipping in joints and feet even after scenes change. The researchers said it could be applied to sports broadcast analysis, digital human production based on films and dramas, and virtual and augmented reality content. However, they left as a follow-up task the problem of continuing to track the same person even in scenes that move to entirely different locations.