A clinical viewer and annotation tool for multi-view intraoral photographs. Standardised five-view photographs are assigned their clinical view, segmented into FDI-numbered tooth instances, and presented in a correction interface where a clinician verifies and fixes the result. Built on human-verified FDI tooth-instance annotations, with baselines for view classification and tooth instance segmentation.
A clinician uploads a patient's photographs. Each one is assigned its clinical view, then segmented into FDI-numbered tooth instances:
photographs → view classification → SegmentAnyTooth (full image) → post-processing → FDI instances
ResNet18 no ROI stage frozen rules
Two design choices in that chain:
- No ROI stage. The viewer segments the full image and crops nothing.
- Post-processing on. Frozen rules clean up the segmenter output; they run after the forward pass in both segmenter paths.
Mask R-CNN is offered as the second segmenter, also on the full image, and is faster. Both are selectable in the viewer.
Download: https://ditto.ing.unimore.it/iop-compass/
Note: for model weights drop us an email at federico.bolelli[at]unimore.it
Four things have to be put in place by hand. This is the layout to end up with —
third_party/ is generated and should never be populated manually:
IOP-Compass/
├── weights/
│ ├── view_classifier.pt ← rename/symlink of classifier__*.pt
│ ├── maskrcnn.pt ← rename/symlink of maskrcnn__*.pt
│ └── SegmentAnyTooth weights/ ← the SAT weight directory, unpacked as-is
│ ├── segmentanytooth_vit_tiny.pt
│ ├── segmentanytooth_yolo11_front.pt
│ ├── segmentanytooth_yolo11_upper.pt
│ ├── segmentanytooth_yolo11_lower.pt
│ └── segmentanytooth_yolo11_right.pt
├── inputs/
│ └── SegmentAnyTooth/ ← the MIT vendor code, git-cloned
│ ├── sam.py
│ └── utils.py
└── third_party/ ← symlinks, created by fetch_third_party.py
├── sat_weights -> ../weights/SegmentAnyTooth weights
├── segmentanytooth_src -> ../inputs/SegmentAnyTooth
└── torch_home
weights/ and inputs/ are the only two directories you touch. Both are
git-ignored, so nothing you put there can be committed by accident.
| file | what it is |
|---|---|
view_classifier.pt |
ResNet18 five-view orientation classifier — assigns each photograph its clinical view |
maskrcnn.pt |
Mask R-CNN R50-FPN tooth-instance segmenter |
They are emailed under the name of the training run that produced them, so rename or symlink them to the two names above — the viewer does no filename guessing and will refuse to start otherwise:
cd weights
ln -s classifier__*.pt view_classifier.pt
ln -s maskrcnn__*.pt maskrcnn.pt
cd ..The code is MIT. It is not redistributed here only because it is upstream's, so
clone it yourself; no agreement and no email are needed. It must land in
inputs/SegmentAnyTooth, with sam.py at the top level of that directory:
git clone https://github.com/thangngoc89/SegmentAnyTooth inputs/SegmentAnyToothThe weights are covered by a separate non-commercial licence — email the
maintainers. They arrive as a directory literally named SegmentAnyTooth weights
(with the space); move it under weights/ without renaming it or flattening it:
mv ~/Downloads/'SegmentAnyTooth weights' weights/python scripts/fetch_third_party.pyThis downloads nothing and accepts no licence. It links the two locations above into
third_party/, records their SHA-256, creates third_party/torch_home, and exits
non-zero naming anything still missing. SAT_CODE_DIR and SAT_WEIGHT_DIR override
the source locations if you keep them elsewhere; cd app && make check then confirms
every file the enabled segmenters need.
The viewer applies no ROI stage, so no ROI-detector, geometric-prior or SAM 3 weights
are needed. SAM 3 is used only by the ROI arm of the benchmark, and
fetch_third_party.py therefore treats it as optional.
python -m venv .venv && source .venv/bin/activate
pip install -r requirements-lock.txt # pins the +cu126 torch index itself
python scripts/fetch_third_party.py # link the SegmentAnyTooth code and weightsThe lock file carries its own --extra-index-url, because the pinned torch and
torchvision wheels have a +cu126 local version and are not on PyPI. It installs
everything the viewer needs, Flask included.
The third step assumes the weights and the vendor code are already in place — see Where everything goes above.
Run the viewer (API on :5000, UI on http://localhost:5173):
cd app && make install # frontend dependencies, once
make check # confirm every weight is present before starting
make dev # backend and UI togethermake check names any missing file and the environment variable that relocates it,
rather than failing on the first clinical image. To serve one segmenter only:
make backend SEGMENTERS=sat.
Run the models on your own images, without the viewer — see
docs/inference.md for the copy-pasteable version:
from iop_compass.segmentation.segmentanytooth_adapter import SegmentAnyToothRunner
from iop_compass.segmentation.postprocessing import PostProcessParams, postprocess_instances
runner = SegmentAnyToothRunner(weight_dir="third_party/sat_weights",
code_dir="third_party/segmentanytooth_src",
device="cuda")
prediction = runner.predict(image_bgr, view_label="frontal")
masks, fdis, scores, _ = postprocess_instances(
prediction.masks, prediction.fdis, prediction.scores,
roi_mask=None, exclusion_mask=None,
params=PostProcessParams.from_yaml("configs/segmentation/postprocessing.yaml"),
view_label="frontal",
)app/ the viewer: Flask backend + React frontend
configs/ dataset, classification and segmentation configuration
docs/ inference and the viewer
internal/ cluster-specific submission tooling, not needed to use the release
scripts/ audit, manifest, splits, training, inference, evaluation, figures
src/iop_compass/
data/ adapter, manifest, rasterisation, validation, splits, imaging
classification/ dataset, model, training, constrained assignment, metrics
roi/ the four ROI strategies, factory, ROI-only evaluation
segmentation/ SegmentAnyTooth adapter, Mask R-CNN, post-processing,
matching, metrics, the benchmark grid, inference
reporting/ aggregation, bootstrap, figures, labels, qualitative panels
tests/ leakage, manifest, label mapping, matching, metrics, grid
results/ and runs/ are produced by a campaign and are not tracked.
| task | variants |
|---|---|
| View classification | C0 pretrained, no augmentation · C1 pretrained + clinically plausible augmentation · C2 from scratch. C1 also evaluated with patient-set constrained assignment. |
| ROI extraction | full image · view-specific geometric prior fitted on training annotations · learned single-box detector · SAM 3 concept prompts with a documented full-image fallback. |
| Tooth instance segmentation | The full ROI × segmenter × post-processing grid: 4 × 2 × 2 = 16 cells, every cell through one code path so the axes are not confounded with a per-backend crop convention. |
The dataset is publicly available at https://ditto.ing.unimore.it/iop-compass/.