Agentic STAR: Shape Tracking and Reconstruction from Monocular Videos

Kirill Mazur, Nikita Karaev, Matthew Chang, Jitendra Malik, Nur Muhammad “Mahi” Shafiullah

1Amazon FAR

our mesh joint axis and pivot (toggle with ⟲) panel: one colour per part

Agentic optimisation timelapse

We demonstrate how the agent iterates on both reconstruction and pose on two selected frames of a video sequence over time.

Two observed frames of the scissors capture, side by side
– / –

Reconstructions on ARCTIC

our mesh GT left hand GT right hand

Reconstructions on HOT3D

our mesh GT left hand GT right hand

BibTeX

@misc{mazur2026agenticstar,
  title        = {Agentic STAR: Shape Tracking and Reconstruction from Monocular Videos},
  author       = {Mazur, Kirill and Karaev, Nikita and Chang, Matthew and Malik, Jitendra and Shafiullah, Nur Muhammad},
  year         = {2026},
  howpublished = {\url{https://agenticstar.github.io/}},
  note         = {Project page}
}