Skip to content
caliperhq

annomate-examples

annomate-examples is the companion corpus for annomate: a research bench of real images — sourced from Wikimedia Commons and other openly-licensed collections — that Claude has annotated through the three-voice perception loop. Each entry ships the original image, the rendered VIA project JSON, and an animated overlay GIF showing the regions Claude landed.

It’s the sister project to jscad-mcp-example. Together they cover both halves of Claude with eyes on the physical world: jscad-mcp-example is what Claude builds; annomate-examples is what Claude sees.

What these are for

Section titled “What these are for”

Not demos — a research bench. Each image was chosen because it tests something specific. The corpus has surfaced a stable set of recurring failure modes:

  • priors override pixels (Hokusai’s Fuji placed on the central foam mass, ~240 px wrong)
  • sub-canvas-tenth features can’t be placed from a thumbnail (jaguar eyes; Vermeer’s brass pail)
  • named-object vs. named-part boxing errors (church = tower only, foot-warmer = pierced cube only)
  • posture-induced face drift on leaning, seated, bent figures (Camus at the Nobel banquet)
  • specialist-vocabulary breaks the VLM verifier (Howe sewing-machine needle terminology)

…and driven a stream of tool changes back into the annomate MCP server itself — fraction coordinates, via_get_image_crop, overlay-on-image, the [ai] extra and its third-voice CLIP/SAM/VLM pipelines.

Section titled “Gallery”

Animated annotation overlays for all 20 sessions. Click any card for the full what-was-annotated writeup.

Jaguar
small-feature placement

Jaguar

Small-feature placement on a contemporary wildlife photograph.

11 regions
Batería de Castillitos
architectural part-naming

Batería de Castillitos

Architectural part-naming on a 1930s coastal artillery site.

10 regions
Mangrove Arch
botanical anatomy

Mangrove Arch

Botanical anatomy of Rhizophora prop-roots and scene context.

11 regions
Häfeli DH-5
specialist vocabulary

Häfeli DH-5

Specialist equipment naming on a 1930 swisstopo aerial photograph.

11 regions
Gunung Lambak Waterfall
geomorphology + flora

Gunung Lambak Waterfall

Geomorphology and flora in a long-exposure tropical stream-scape.

11 regions
Marmalade Hoverfly
entomology

Marmalade Hoverfly

Entomology at near-collision intercept — Episyrphus balteatus anatomy.

11 regions
Takht-i-Bahi Monastery
aerial interpretation

Takht-i-Bahi Monastery

Aerial drone view of a 1st-century CE Gandharan Buddhist monastery.

15 regions
GRACE-FO Launch
prior reconciliation

GRACE-FO Launch

Aerospace part-naming with metadata prior reconciliation.

14 regions
Marché Dauphine
mixed-shape vocabulary

Marché Dauphine

Mixed-shape vocabulary — circles, ellipse, polygons, and rectangles.

17 regions
Deposition Fresco
iconographic / figural

Deposition Fresco

Iconographic figural labelling — Paolo Caylina the Younger, c. 1500.

13 regions
Avercamp, Winter Landscape
dense scene

Avercamp, Winter Landscape

Dense genre scene — one region per activity-type, not per figure.

23 regions
Vermeer, The Milkmaid
spatial-prior failure

Vermeer, The Milkmaid

"Directly below" spatial-prior failure — brass pail 440 px off on first pass.

19 regions
Rembrandt, The Night Watch
polyline-identity error

Rembrandt, The Night Watch

Polyline-identity error — calf line read as a plausible polearm shaft.

17 regions
Hokusai, The Great Wave
iconographic prior

Hokusai, The Great Wave

Iconographic prior — Fuji placed on the central foam mass, ~240 px wrong.

13 regions
Van Gogh, Self-Portrait (1887)
technique vocabulary

Van Gogh, Self-Portrait (1887)

Technique vocabulary — radial dabs, divisionism, impasto as annotation targets.

14 regions
Van Gogh, Hill at Montmajour
mark vocabulary

Van Gogh, Hill at Montmajour

Mark vocabulary — stipple, hatch, contour, and bare paper as annotation targets.

13 regions
Lupa Capitolina
multi-era object

Lupa Capitolina

Three historical eras in one object — Etruscan wolf, Renaissance twins, modern mount.

9 regions
Camus at the Nobel Banquet, 1957
posture-induced drift

Camus at the Nobel Banquet, 1957

Posture-induced face drift on three leaning figures.

13 regions
Camus Crowns the Lucia, 1957
profile-face drift

Camus Crowns the Lucia, 1957

Profile-face drift and standing-vs-seated prior failure.

10 regions
Howe's 1846 Sewing Machine
cross-view correspondence

Howe's 1846 Sewing Machine

Cross-view correspondence — the curved eye-pointed needle identified across two views.

18 regions

Browse the full gallery →

Layout

Section titled “Layout”
examples/
  NN-<example-name>/
    README.md            attribution, license, source URL, what was annotated
    <image>.{jpg,png}    original from Wikimedia Commons
    annotations.json     VIA v3 project file (load via "Project → Open" or via_update_project)
    annotated.gif        animated overlay preview

Quick start

Section titled “Quick start”
Terminal window
git clone https://github.com/caliperhq/annomate-examples.git
cd annomate-examples

Open any example’s annotations.json in VIA to see the regions:

  1. Install and start annomate (pip install annomate && annomate --browser)
  2. In the browser UI: Project → Open → pick the annotations.json

Or ask Claude to load it via the MCP tools:

via_update_project(<contents of annotations.json>)

Image attributions

Section titled “Image attributions”

Every image is sourced from Wikimedia Commons (or another openly-licensed source). Each example’s README.md carries the source URL, the original author / uploader, the license (CC-BY-SA, CC-BY, public domain, etc.) with its version, and any attribution text the license requires. The annotations and code in this repo are MIT-licensed; the images themselves keep their original license.