Auto-rigging children's drawings on a phone without a learned model: thinning, skeleton graph, geodesic skin weights
retained text
I wanted Meta's Animated Drawings (Smith et al. 2023) working offline on an iPhone, with no server and no model weights, so I rebuilt the pipeline from classic parts. The cutout takes 0.1–0.3 s in my tests. The video shows each step on one drawing. **Mask**: Ink is anything darker than the local paper (box blur at about 1/18 of the image size, luma difference > 12) or clearly more colourful than it (chroma). Close with a disc about 1/90 of the image size, flood-fill the paper from the border, and whatever the flood can't reach is the figure. A second, bigger closing seals faces drawn with a faint outline, but only when the enclosed region is coloured in, so the paper between an arm and the body isn't filled. **Skeleton**: Max-pool the mask to about 140 px, run Zhang–Suen thinning, and prune spurs relative to the local thickness (a short branch off a thick trunk is an outline bump; the same length on a thin limb is a finger). Ends and junctions come from the crossing number. Head = the thickest point near the top, feet = the lowest ends on different legs, hands = ends that join the trunk below the neck and above the knees, shoulders and hips where the limb enters the body. Missing limbs collapse instead of failing, so a one-armed monster or a legless princess still rigs. **Skinning**: A triangle grid of about 44 cells. Per-bone weights use geodesic distance \*inside\* the (bridged) mask, 1/(d+1)\^4, keeping the top 4. That stops two legs drawn touching from moving together, which plain Euclidean weights get wrong. Linear blend skinning in Metal. **Photos** of real people use Vision's body pose instead; photos of things use Vision's subject lift for the mask. On 8 test drawings the rig was right on 7. The miss is a dinosaur whose back spike becomes its right hand. The open problem is limbs drawn overlapping the body: the thinning can't see them. Has anyone had luck with a cheap heuristic for that, short of a learned pose model? The app is free (iPhone/iPad), if you want to try your own drawings: [https://apps.apple.com/app/id6820002786](https://apps.apple.com/app/id6820002786)
document fetched 2026-10-13 00:24 · canonical https://reddit.com/r/computervision/comments/1x3vrz5/autorigging_childrens_drawings_on_a_phone_without