| CARVIEW |
Select Language
HTTP/2 200
server: GitHub.com
content-type: text/html; charset=utf-8
last-modified: Wed, 26 Nov 2025 01:08:46 GMT
access-control-allow-origin: *
etag: W/"6926531e-3f4b"
expires: Tue, 30 Dec 2025 05:15:27 GMT
cache-control: max-age=600
content-encoding: gzip
x-proxy-cache: MISS
x-github-request-id: D4A1:234FE9:9B3AA4:AE3E4F:69535D97
accept-ranges: bytes
age: 0
date: Tue, 30 Dec 2025 05:05:27 GMT
via: 1.1 varnish
x-served-by: cache-bom-vanm7210067-BOM
x-cache: MISS
x-cache-hits: 0
x-timer: S1767071127.375701,VS0,VE211
vary: Accept-Encoding
x-fastly-request-id: e4637b2653c747b3c153a2a408d99ae5a51b2125
content-length: 3863
FlowFeat: Pixel-Dense Embedding of Motion Profiles
FlowFeat:
FlowFeat:
Pixel-Dense Embedding of Motion Profiles
✨ NeurIPS 2025 Spotlight ✨
The gist:
Distilling optical flow networks into pixel-level task-agnostic representations.
Input
DINOv2
V-JEPA
VideoMAE
FeatUp
FlowFeat
FlowFeat provides versatile pixel-level features. Using motion-driven embedding statistics, it achieves high spatial precision and temporal consistency. Shown: PCA visualization of feature maps vs. state-of-the-art encoders.
More FlowFeat Examples
Hover to view
Hover to view
Hover to view
Hover to view
Hover to view
Hover to view
Video Object Segmentation
FlowFeat is a fine-grained feature representation, where dynamic objects are particularly prominent.
Comparison to FeatUp (Fu et al., 2024)
FlowFeat scales well with the input resolution – here, increased by a factor of 2.
Comparison to LoftUp (Huang et al., 2025)
FlowFeat fares well against supervised models such as LoftUp and can be trained in a completely unsupervised fashion.
Citation
@inproceedings{Araslanov:2025:FlowFeat,
author = {Araslanov, Nikita and Sonnweber, Anna and Cremers, Daniel},
title = {{FlowFeat}: Pixel-Dense Embedding of Motion Profiles},
booktitle = {NeurIPS},
year = {2025},
}
This site uses cookies for analytics.