The authors froze a self-supervised DINOv3 vision transformer and used its patch-level tokens as the acquisition space for deciding which pixels should be labeled.
Selecting ViT tokens cuts labels needed for semantic segmentation
TEAL combines diversity sampling in a transformer’s native token space with pixel-level uncertainty, improving segmentation from as little as one labeled pixel per image.
Big Tech
Utku Cicek · Mahip Singh · Ziyao Shang · Kimathi Kaai · C Thomas · Pedro Pablo Guerrero Vela · +2 more
University of Waterloo · Apple
Research Digest··2 min read
Thread:Visual Token Efficiency
Cicek and colleagues present TEAL, an active-learning method for training semantic segmentation models with extremely sparse pixel annotations.
Why this paper
From Apple and University of Waterloo · Part of Visual Token Efficiency, now 6 papers
In one line
Pixel-level active learning for semantic segmentation works best when candidate selection happens in ViT token space, not interpolated pixel space.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§