AllTracker: Efficient Dense Point Tracking at High Resolution
Published in ICCV, 2025
We introduce AllTracker: a model that estimates long-range point tracks by computing flow fields between a query frame and every other frame of a video. Unlike existing methods, our approach delivers high-resolution, dense correspondence fields that can track hundreds of frames at once. The model uses an efficient architecture with iterative inference on low-resolution grids, combining 2D convolutions for spatial propagation and pixel-aligned attention for temporal propagation. With only 16 million parameters, it achieves state-of-the-art point tracking accuracy at high resolution (768x1024 pixels) and can be trained on diverse datasets for optimal performance.
Recommended citation: Adam W. Harley, Yang You, Xinglong Sun, Yang Zheng, Nikhil Raghuraman, Yunqi Gu, Sheldon Liang, Wen-Hsuan Chu, Achal Dave, Pavel Tokmakov, Suya You, Rares Ambrus, Katerina Fragkiadaki, Leonidas J. Guibas. (2025). AllTracker: Efficient Dense Point Tracking at High Resolution. ICCV 2025.