#berkeleyedu #githubcom #nrenavymil
The research paper “**Humans in 4D: Reconstructing and Tracking Humans with Transformers**” from the University of California, Berkeley, introduces a method for 3D reconstruction of humans from any video and tracking them over time. The research utilizes a network called HMR 2.0, which is a fully “transformerized” version designed for human mesh recovery. This network has enhanced capabilities, particularly in analyzing unusual human poses that were previously challenging to reconstruct from single images. When analyzing videos, the 3D reconstructions from HMR 2.0 feed into a tracking system operating in 3D, which is adept at handling multiple people and maintaining their identities even during occlusions. The resulting system, named 4DHumans, sets new standards for tracking humans from single-view videos. The research also shows that HMR 2.0 is effective in action recognition tasks, surpassing earlier pose-based action recognition techniques. Comparative results indicate HMR 2.0’s superior stability and alignment over its predecessors. This study received support from various agencies and programs, including DARPA and ONR MURI.
=> https://shubham-goel.github.io/4dhumans/