This is a research codebase from Google DeepMind for Tracking Any Point (TAP) in videos. It implements multiple model generations—TAP-Net, TAPIR, BootsTAPIR, TAPNext, TAPNext++, and TRAJAN—along with accompanying benchmarks (TAP-Vid, TAPVid-3D, RoboTAP). The repository provides training pipelines (Jaxline-based), evaluation scripts, interactive colab demos, and a VOT2026 tracker submission. While its primary character is data/research (focused on model design, dataset generation, and metric computation), it also includes application-like live demos that connect to webcams for real-time point tracking. The codebase is well-structured with separate packages for models, utilities, configurations, and datasets, and it supports both JAX/Haiku and PyTorch frameworks. The repository's strength lies in its modular utilities—coordinate transforms, loss functions, parallel scan operations, and augmentation routines—that are cleanly abstracted and could be extracted for use in other video understanding projects. However, it carries typical research-code baggage: dual framework implementations increase maintenance burden, some scripts have hardcoded paths, and full reproduction of benchmarks requires access to licensed datasets (Aria, Waymo, Panoptic Studio) or manual downloads. The codebase is actively developed (last commit dated 2026) and uses modern libraries, but the reliance on custom dependencies like kubric and recurrentgemma may pose compatibility risks.