<?xml version="1.0"?>
<metadata xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:dc="http://purl.org/dc/elements/1.1/"><dc:title>A memory-based deep learning method for long-term visual point tracking</dc:title><dc:creator>Đoković,	Lazar	(Avtor)
	</dc:creator><dc:creator>Kristan,	Matej	(Mentor)
	</dc:creator><dc:creator>Lukežič,	Alan	(Komentor)
	</dc:creator><dc:subject>Computer vision</dc:subject><dc:subject>Deep learning</dc:subject><dc:subject>Visual point tracking</dc:subject><dc:subject>Correspondence</dc:subject><dc:subject>Optical flow</dc:subject><dc:description>Dense point tracking estimates long-range trajectories for all reference pixels, making it a powerful representation of video motion. This is particularly challenging in a causal online setting, where future frames are unavailable at inference time, while points may exhibit substantial motion, and undergo occlusion and appearance changes. Maintaining reliable point identities over long videos thus remains challenging, especially when appearances must be updated over time without corrupting dense and temporally stable predictions. We introduce DeMeTer, a dense online point tracker that combines memory-conditioned global matching with localized motion refinement. Instead of matching only static reference features, DeMeTer maintains a confidence-gated appearance memory that updates each point with reliable past observations, enabling re-detection after occlusion or appearance changes. Simultaneously, a separate motion memory stores recent local dynamics, allowing the refininig module to generate accurate and temporally consistent trajectories. This reference-aligned dual-memory architecture separates identity preservation from short-term motion modeling, making explicit re-detection feasible at dense scale while remaining fully causal. DeMeTer establishes a new state-of-the-art by outperforming related methods by 4.1\% in $\delta_{\mathrm{avg}}$, while comfortably running in realtime.</dc:description><dc:date>2026</dc:date><dc:date>2026-10-22 14:00:07</dc:date><dc:type>Magistrsko delo</dc:type><dc:identifier>190102</dc:identifier><dc:identifier>VisID: 38509</dc:identifier><dc:language>sl</dc:language></metadata>
