google-deepmind/tapnet
Tracking Any Point (TAP)
What it solves
This project provides a suite of tools and models for "Tracking Any Point" (TAP), which allows for the precise tracking of any trackable point on a solid physical surface across video frames. Unlike traditional box or segment tracking, TAP focuses on point-level precision, handling deformable surfaces and long-term tracking across diverse objects without requiring class-specific training.
How it works
The repository implements several evolving architectures:
- TAPIR: A two-stage algorithm that first independently locates candidate point matches for a query point on every frame, then refines the trajectory and query features using local correlations.
- TAPNext / TAPNext++: Formulates point tracking as a next-token prediction problem, propagating information through a network to track points. TAPNext++ specifically improves stability for long-term tracking, occlusions, and re-detection.
- BootsTAP: A training method that uses large unlabeled real-world video datasets to improve accuracy by enforcing consistency across spatial transformations and corruptions.
- TRAJAN: A point trajectory autoencoder that reconstructs missing point trajectories based on a set of support trajectories, allowing for the comparison of motion patterns across different videos.
Who it’s for
- Computer Vision Researchers: Those working on point tracking, optical flow, and motion analysis.
- Robotics Engineers: Developers using RoboTAP to execute manipulation tasks through visual imitation.
- AI Developers: Anyone needing high-precision point tracking for video analysis or generative video evaluation.
Highlights
- Multiple Benchmarks: Includes TAP-Vid, RoboTAP, and TAPVid-3D for 2D and 3D point tracking evaluation.
- Diverse Model Options: Offers various checkpoints including TAPIR, BootsTAPIR, and TAPNext, with implementations in both Jax and PyTorch.
- Real-time Capability: Includes a causal TAPIR model capable of real-time tracking on GPU platforms.
- Robotics Integration: Specifically extends point tracking for real-world robotics manipulation via RoboTAP.