sv

Everything you need to build a Svelte project, powered by sv.

Creating a project

If you're seeing this, you've probably already done this step. Congrats!

# create a new project
npx sv create my-app

To recreate this project with the same configuration:

# recreate this project
npx [email protected] create --template minimal --types ts --install npm OmniPlot

Developing

Once you've created a project and installed dependencies with npm install (or pnpm install or yarn), start a development server:

npm run dev

# or start the server and open the app in a new browser tab
npm run dev -- --open

Building

To create a production version of your app:

npm run build

You can preview the production build with npm run preview.

To deploy your app, you may need to install an adapter for your target environment.

---

Trace Image Algorithm (src/lib/utils/trace.ts)

Converts a raster PNG/JPG of a window/PPF template into a closed SVG path.

Pipeline (v4)

1. Otsu threshold — binarize RGBA image to foreground mask
2. Moore contour tracing — extract closed pixel boundary (~1–2 px staircase)
3. Light Gaussian smooth (σ=1.5) — damp staircase noise while keeping corner coords within ~0.5 px of true geometry
4. detectCornersFromCurvature — Menger curvature on light-smooth contour; threshold = mean + 1.5σ; NMS within minSpacingFrac=6% of contour length
5. normalizePts — map pixel contour to 0–100 path-unit space (preserves aspect ratio, 3 px margin)
6. refineCornerPositions (maxMove=0.5) — PCA line-fit + parametric intersection snap; bounded to 0.5 units so it only corrects genuine sub-pixel notch artifacts (Moore produces a ~0.5 px notch at true corners); snaps > 0.5 are rejected
7. buildPathFromCorners — Schneider cubic bezier fitting per span, G1 tangent continuity at smooth junctions, tolerance 2.0 px in path units
8. Hausdorff gate — if max deviation ≤ 2.0 path-units, return; otherwise retry with tighter tolerance (fallback pass)

Current metrics (fixture: src/lib/utils/__fixtures__/car-window-ref.png, 152×109 px)

Metric	Value
IoU vs truth mask	98.4%
Hausdorff deviation	1.308 path-units
Corners detected	4 (structural transitions only)
Bezier segments	11 C commands
Tests	63/63 passing

Known limitations / next improvement areas

• Single test fixture — the committed fixture is a clean synthetic shape. Real photos introduce JPEG noise, anti-aliasing, and gradual tonal transitions that stress the threshold and contour steps. More real-world test images are needed to push the algorithm further.
• mean + 1.5σ threshold — susceptible to outlier inflation on shapes with one dramatically sharper corner among several gentler ones. median + 3×MAD was tested but collapses to ~0 on mostly-flat contours (straight sides dominate the κ distribution, making median ≈ MAD ≈ 0). A robust alternative (e.g. trimmed mean, or percentile-based floor) is needed.
• refineCornerPositions rarely fires — with maxMove=0.5, no corner on the current fixture qualifies (smallest actual snap distance was 1.1 units). The Moore notch artifact is real (~0.5 px) but the line-fit intersection still overshoots on curved-edge windows. A better approach might be direct subpixel gradient-following rather than PCA intersection.
• Segment count — 11 bezier segments for a 4-corner window is reasonable but not minimal. Reducing segments without losing shape fidelity (especially the A-pillar inflection) requires better span-level curvature estimation.