DermDepth: Monocular Metric Scale 3D Reconstruction for Dermatology

Model checkpoints from DermDepth (Carrión & Norouzi, MICCAI 2026).

Checkpoints

Filename	Training data	Notes
DermDepth_Synth.pt	D-Synth only	"DermDepth_S" in the paper. Synthetic-only baseline.
DermDepth_Synth_SKINL2_WoundsDB.pt	D-Synth → SKINL2 + WoundsDB	Intermediate stage (before DDI pseudo-GT).
DermDepth_Synth_SKINL2_WoundsDB_DDI.pt	D-Synth → SKINL2 + WoundsDB → DDI pseudo-GT	Best model. Corresponds to "DermDepth" in the paper.
DermDepth_Synth_Normals.pt	D-Synth, normal-head emphasis	Trained normal-head model.

Key results (held-out test sets, paper Table 1)

Method	SKINL2 Scale	WoundsDB Scale	DDI Ratio	Fitzpatrick Disparity
MoGe-2 (baseline)	16.10×	0.62×	81.0×	10.90
DermDepth_Synth	1.11×	0.28×	9.2×	1.70
DermDepth (best)	0.87×	0.91×	1.95×	1.02

Scale ratio target is 1.0×. See the paper for full benchmarks and SI-δ₁ details.

Usage

These checkpoints are designed to be loaded by MoGe-2, modified per the DermDepth code repository. The repo contains end-to-end training, inference, and evaluation scripts.

Quick download:

from huggingface_hub import hf_hub_download
ckpt = hf_hub_download(
    repo_id="hcarrion/DermDepth",
    filename="DermDepth_Synth_SKINL2_WoundsDB_DDI.pt",
)

Citation

@inproceedings{carrion2026dermdepth,
  title     = {DermDepth: Toward Monocular Metric Scale 3D Reconstruction Models for Dermatology},
  author    = {Carri{\'o}n, H{\'e}ctor and Norouzi, Narges},
  booktitle = {Medical Image Computing and Computer-Assisted Intervention (MICCAI)},
  year      = {2026}
}

License

CC BY-NC 4.0 (research / non-commercial use). The base MoGe-2 weights remain under their original license — see the MoGe repository for details.

Related Resources

• Code & training scripts: https://github.com/hectorcarrion/dermdepth
• D-Synth dataset (training data): https://huggingface.co/datasets/hcarrion/D-Synth
• Base model: https://huggingface.co/Ruicheng/moge-2-vitl-normal