SlotMatch: Distilling Temporally Consistent Object-Centric Representations for Unsupervised Video Segmentation

Diana-Nicoleta Grigore, Neelu Madan, Andreas Møgelmose, Thomas B. Moeslund, Radu Tudor Ionescu · 2025

Read SlotMatch: Distilling Temporally Consistent Object-Centric Representations for Unsupervised Video Segmentation by Diana-Nicoleta Grigore, Neelu Madan, Andreas Møgelmose, Thomas B. Moeslund, Radu Tudor Ionescu (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-2508-03411.txt.

Description

Diana-Nicoleta Grigore and co-authors present SlotMatch in 2025, a lean way to segment video without labels. Sorting objects in video with no supervision is hard, and leading slot-based models are large and costly. The team wanted the same object sense in a small student model.

SlotMatch teaches the student by aligning its object slots with a teacher's slots, using direction of the slot vectors only. Size of the vectors, which can encode false confidence, is ignored. No extra losses or labels are needed. The simple scheme transfers object structure at low cost. Such light models could run video sorting on modest hardware.

Catalog record

CREATORDiana-Nicoleta Grigore, Neelu Madan, Andreas Møgelmose, Thomas B. Moeslund, Radu Tudor Ionescu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-03411

Browse the full library · About the Tower of Babel