🦷 3D Teeth Landmark Detection

A Point-Transformer-v3-style network that detects six classes of anatomical dental landmarks directly on 3D intraoral scan point clouds — no tooth segmentation stage. Built during a research internship at the National Dental Centre of Singapore and evaluated under the official 3DTeethLand (MICCAI 2024) protocol.

Model & weights ↗ Code ↗ 3DTeethLand challenge ↗ 6.45M params · ~0.9 s / scan

Model output on a held-in example scan

Arch

Show Predicted Ground truth Scan surface

Drag to rotate · scroll to zoom · click legend entries to isolate a landmark class. Enable “Ground truth” (hollow diamonds) to compare against the clinician annotations.

Accuracy on this scan

Class	Pred	GT	Median err	< 2 mm

Error = distance from each predicted landmark to the nearest ground-truth landmark of the same class.

Official challenge results

Arch	mAP	mAR
Lower	0.536	0.412
Upper	0.553	0.438

Scored with the challenge's own evaluation code on the 50-scan held-out test split per arch.

Tolerance	0.5 mm	1.0 mm	1.5 mm	2.0 mm
mean AP	0.091	0.473	0.701	0.790
mean AR	0.263	0.635	0.787	0.843

Lower arch, test split.

How it works

Input — 10,000 vertices sampled from the scan mesh, each with its surface normal: [N, 6] features (x, y, z, nx, ny, nz).
Encoder — 4 stages × 2 pre-norm transformer blocks (multi-head self-attention + MLP) at hidden dim 256. Unlike standard PTv3, no downsampling is applied, so per-point resolution survives end to end.
Distance-map regression — a per-point MLP predicts, for each of the six classes, the normalized distance to the nearest landmark of that class (clamped at 15 mm, sqrt-sharpened). Dense targets give every point a gradient — this trained far better than the sparse Gaussian heatmaps tried first.
Post-processing — points below a per-class threshold are clustered with DBSCAN; each cluster yields one landmark. Handles variable landmark counts (missing teeth) without knowing the number in advance.