Vizualno sledenje točk s pomočjo temeljnega 3D-modela sveta

In this master's thesis, we explore the use of 3D world foundational models for the purpose of visual object tracking, specifically for the task of tracking arbitrary points across video sequences. We develop a method that leverages the pixel-wise feature extracting capabilities of the MASt3R architecture. When cross-correlated, these features provide robust estimates of the point's new location in subsequent frames. To propagate point locations across multiple frames, we employ a chaining mechanism based on the concept of anchor frames. To refine point predictions and estimate occlusion probabilities, we utilize a transformer-based refinement module inspired by LocoTrack. The method is trained on synthetic point tracking datasets, and evaluated on the TAP-Vid benchmark, where it achieves performance comparable to the state-of-the-art. On a subset of the first 100 TAP-Vid Kinetics videos, our method achieves the highest position accuracy of 66.3, surpassing LocoTrack by 0.9 and CoTracker3 by 1.5, albeit at a significantly lower processing speed.