Back to browse
Lingbot Vision – Self-Supervised Dense Perception (0.296 NYUv2 RMSE)

Lingbot Vision – Self-Supervised Dense Perception (0.296 NYUv2 RMSE)

by Kajaking·Jul 7, 2026·1 point·0 comments

AI Analysis

●●●BangerBig BrainWizardry

Beats 7B DINOv3 on depth estimation with 1B params using boundary-centric self-supervision.

Strengths
  • Masked boundary modeling learns sub-pixel boundaries without annotations or edge detectors.
  • 0.296 NYUv2 RMSE beats DINOv3's 0.309 despite being 7x smaller parameter count.
  • Frozen features scale to 4096-token grids preserving whiskers and thin structures.
Weaknesses
  • Only 55 GitHub stars suggests limited community adoption or recent release.
  • Dense perception niche may not appeal to general ML practitioners.
Category
Target Audience

ML researchers, computer vision engineers, robotics developers

Similar To

DINOv3 · Segment Anything (SAM) · Depth Anything

Similar Projects