DELTAv2: Accelerating Dense 3D Tracking

Tuan Duc NgoUMass AmherstAshkan MirzaeiSnap Inc.Guocheng QianSnap Inc, Hanwen LiangUniversity of TorontoChuang GanUMass AmherstEvangelos KalogerakisUMass Amherst & TU CretePeter WonkaSnap Inc. & KAUSTChaoyang WangSnap Inc · 2025

Read DELTAv2: Accelerating Dense 3D Tracking by Tuan Duc NgoUMass AmherstAshkan MirzaeiSnap Inc.Guocheng QianSnap Inc, Hanwen LiangUniversity of TorontoChuang GanUMass AmherstEvangelos KalogerakisUMass Amherst & TU CretePeter WonkaSnap Inc. & KAUSTChaoyang WangSnap Inc (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-01170.txt.

Description

Tuan Duc Ngo and co-authors present a faster method for tracking many points in three dimensions across a video, in a paper from 2025. Dense tracking is used in effects, robotics, and scene analysis. The best past methods are slow when they follow a huge number of paths at once.

The team found two main bottlenecks. Repeated updates with a Transformer model grow costly as paths pile up, and building the match features for each point is also heavy. Their fixes start with a small set of points and expand step by step, seeding new paths with a learned filler module. A second fix cuts the cost of the match features. Together, the changes gave speed gains from five-fold to a hundred-fold while keeping top accuracy.

Catalog record

CREATORTuan Duc NgoUMass AmherstAshkan MirzaeiSnap Inc.Guocheng QianSnap Inc, Hanwen LiangUniversity of TorontoChuang GanUMass AmherstEvangelos KalogerakisUMass Amherst & TU CretePeter WonkaSnap Inc. & KAUSTChaoyang WangSnap Inc
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-01170

Browse the full library · About the Tower of Babel