TAPVid-MV

Overview

The first benchmark for 3D point tracking across several moving cameras.Track a query point through synchronized views, in one shared world frame.
284 sequences, 1,142 calibrated camera streams, 109,769 3D tracks.Seven subsets, from robot manipulation and driving to two-person interaction and synthetic scenes.
Ground truth you can easily view yourself.Every sequence is manually verified and published as an interactive 3D recording.
Geometry and correspondence, scored separately.We evaluate reconstruction quality and tracking accuracy on the same sequences.
Over 30 baselines under one protocol, but the task remains unsolved.No method comes close, and multi-view trackers do not consistently beat monocular ones.
Training data included.The Perpetua generator and metric 3D trajectories for 5,371 robot episodes ship with the benchmark.

The seven TAPVid-MV subsets — DROID, EgoExo4D, Harmony4D, PACE, Hi4D, Waymo, and Perpetua — each shown as three camera views of one sequence with its camera-motion glyph and size. — The seven subsets. Each column shows three views of one sequence with its ground-truth tracks, and below them the camera trajectories of that sequence (white dots are static cameras, blue paths moving ones), how many of its cameras move, how far they travel, and the size of the subset.

Inspect any sequence in 3D

Every one of the 284 sequences is released as a Rerun recording — multi-view RGB, colored point cloud, ground-truth 3D tracks, and camera frusta. Pick one below, or browse the full index.

Subset Sequence

Choose a sequence and press Load the viewer streams the recording from the release bucket

BibTeX

@article{tapvidmv2026,
  title   = {TAPVid-MV: A Benchmark for Tracking Any Point in 3D Across Multiple Views},
  author  = {Skanda Koppula and Frano Raji{\v{c}} and Abdullah Faiz Ur Rahman and Yi Yang and Ignacio Rocco and Jeet Thakwani and Rishabh Kabra and Andrew Zisserman and Joao Carreira and Siyu Tang and Carl Doersch and Gabriel Brostow},
  year    = {2026}
}

Citing the source datasets

TAPVid-MV is built on six existing datasets, whose authors made this benchmark possible. If you use a subset, please cite the dataset it comes from alongside this work:

DROID robot manipulation
Ego-Exo4D sporting activities
Harmony4D close human interaction
PACE tabletop objects
Hi4D two-person interaction
Waymo Open Dataset driving