Polarini Focusini

Automated in-focus region detector that blends monocular depth estimation with classic frequency-domain (Difference of Gaussian - DoG) sharpness cues.

Example

What happens under the hood

1. Runs Depth-Anything V2 to get floating-point depth.
2. Builds a 3-level Gaussian pyramid and two Difference-of-Gaussians (DoG) maps.
3. Applies Non-Maximum Suppression in space and across scales.
4. Keeps only strong extrema → votes for focus → finds dominant depth bins.
5. Saves every intermediate step to a per-image debug/ folder for easy inspection.

polarini_focusini.mp4

Live coding walkthrough 🎬

Watch the 90-minute live “vibe-coding” session that produced this repo (English code + English subs + Russian-language commentary):

Please cite ⭐

@misc{poliarnyi2025,
  title        = {Polarini Focusini: open-source pipeline for in-focus region detection},
  howpublished = {\url{https://github.com/UnicornGlade/PolariniFocusini}},
  author       = {Poliarnyi, N.},
  year         = {2025},
  note         = {YouTube demo: “Finding Focus in Photos Using Depth Anything and DoG”}
}

Stars, forks, issues – all very welcome!