tapnet

Why we included this project

Tracking Any Point (TAP) is the problem of following a single physical point as it moves through a video, and this Google DeepMind repository holds the reference implementations for that task. It covers several generations of models: TAPIR, the two-stage matcher-and-refiner that set the standard on the TAP-Vid benchmark; BootsTAPIR, which learns consistency on unlabeled real footage and beats its predecessor by a clear margin; and TAPNext, the newest and simplest tracker, which frames the problem as token prediction. The same repo ships the TAP-Vid and TAPVid-3D benchmarks with ground-truth trajectories, so you can score your own model against a fixed standard instead of improvising an evaluation. Robotics teams will find RoboTAP useful, since it turns those point tracks into imitation-based manipulation in the real world. With model code, evaluation scripts, and pretrained weights included, it works as a research baseline or as a starting point for a custom tracker.