Common Pets in 3D: Dynamic New-View Synthesis of Real-Life Deformable Categories

Reconstruct unseen videos at test-time Train a category-level model on videos of non-rigid objects TrackeRF TrackeRF Figure 1. We tackle the problem of synthesising new views of deformable objects given only a small number of views taken at different times. We introduce new benchmark data for this task: Common Pets in 3D (CoP3D), containing 4,200 smartphone videos of cats and dogs collected 'in the wild'. We also propose a new method, Tracker-NeRF, a deformable new-view synthesis algorithm which learns a category-level reconstruction prior from videos and applies it to reconstruct new objects at test time.