ayoubft/fraXteX

Models — U-Net, SegFormer and SAM 2 for Automated Fracture Detection, trained on FraXet

https://doi.org/10.5281/zenodo.17866853

Model Details

License: MIT Dataset: FraXet (zenodo) Repository: github.com/ayoubft/fractex2D.pt Demo: huggingface.co/spaces/ayoubft/fractex2D_tuto Paper: doi.org/10.5194/egusphere-2026-1097 (preprint)

These models perform pixel-wise fracture segmentation on outcrop imagery. They serve as baseline architectures in the FraXet benchmarking framework, which compares classical filters, CNNs and transformer models for geological fracture mapping.

Two input modalities are published, and this is a deliberate axis of the benchmark:

• rgbdem/ — 4-channel models taking paired RGB + DEM patches.
• rgb/ — 3-channel models taking RGB only, for settings where no DEM is available.

Models

File	Arch.	Modality	In	Backbone	Params	Tile	Preprocessing
rgbdem/unet-rgbdem.onnx	U-Net [1]	RGB + DEM	4	from scratch, init_features=64	31.03 M	256×256 fixed	external
rgbdem/segformer-rgbdem.onnx	SegFormer [2]	RGB + DEM	4	smp decoder on ResNet-34 (ImageNet)	21.87 M	256×256 fixed	external
rgb/unet-rgb.onnx	U-Net [1]	RGB	3	from scratch, init_features=64	31.03 M	any multiple of 16	in-graph
rgb/sam2-rgb.onnx	SAM 2 [3]	RGB	3	SAM 2.0 hiera-tiny, encoder frozen	33.12 M total	512×512 fixed	in-graph

Every model emits a single-channel sigmoid map: per-pixel fracture probability in [0, 1].

Each .onnx file carries its own input contract as embedded metadata, so the contract travels with the file however you obtain it:

import onnx
md = {p.key: p.value for p in onnx.load("rgb/unet-rgb.onnx").metadata_props}
# {'model.name': 'unet', 'input.channels': '3', 'input.range': '0-255',
#  'preprocessing.in_graph': 'true', 'preprocessing.normalize': 'dataset',
#  'citation.doi': '10.5194/egusphere-2026-1097', ...}

All four models expose the same 15 keys, including citation.doi, so a file separated from this model card still says what to cite. There are no separate metadata files.

SAM 2 note. This is not SAM 2 used off the shelf. The hiera-tiny image encoder is frozen; only the mask decoder was fine-tuned on FraXet. The graph is prompt-free: it takes an image and returns a mask, with no prompt encoder in the export. The 33.12 M figure is the total exported parameter count, most of which is the frozen encoder.

Input contracts

The two modality directories do not share a preprocessing path. Getting this wrong produces plausible-looking but wrong masks rather than an error.

	rgbdem/ (4-channel)	rgb/ (3-channel)
Feed as	float32 scaled to [0, 1]	float32 holding raw 0–255 values
Normalization	none applied; DEM band normalized separately	baked into the graph
Exported with	torch 2.7.1, TorchScript exporter	torch 2.13, dynamo exporter
Opset	14–15	17–18

The rgb/ models carry their own mean/std as graph constants, so you must not normalize before calling them — just hand over raw pixel values as float32.

The two rgb/ models use different constants, deliberately:

• unet-rgb uses FraXet dataset statistics (mean 0.6064, 0.5927, 0.5793, std 0.1688, 0.1622, 0.1621), because it is trained from scratch on FraXet.
• sam2-rgb uses ImageNet statistics (mean 0.485, 0.456, 0.406, std 0.229, 0.224, 0.225), because its frozen encoder was pretrained under that distribution and should keep seeing it.

Since both sets are inside their graphs, this asymmetry costs the caller nothing — but it is a real difference between the two arms and is stated here rather than left to be discovered.

Usage

import numpy as np, onnxruntime as ort
from huggingface_hub import hf_hub_download

path = hf_hub_download("ayoubft/fraXteX", "rgb/unet-rgb.onnx")
sess = ort.InferenceSession(path)

# raw 0-255 RGB, NCHW, float32 — no normalization: it is in the graph
x = rgb_uint8.transpose(2, 0, 1)[None].astype(np.float32)
prob = sess.run(None, {sess.get_inputs()[0].name: x})[0]   # (1, 1, H, W) in [0,1]

For rgbdem/, stack the DEM as a 4th band, scale to [0, 1], and tile into non-overlapping 256×256 patches.

Uses

Direct use: predict fracture probability maps or binary masks for UAV or field imagery. Downstream use: baseline models, or assistive pre-annotation tools for geoscience datasets.

Bias, Risks, and Limitations

Predictions depend on annotation quality, illumination and lithology. Thin or poorly illuminated fractures may be missed; shadows and texture can yield false positives. Treat predictions as assistive probability maps and validate with expert interpretation.

The rgbdem/ and rgb/ arms were exported at different times with different toolchains, so a difference in their outputs is not purely architectural.

Repository history

This repo is ONNX-only. It previously shipped PyTorch checkpoints at pytorch/unet.pt and pytorch/segformer.pt, alongside an onnx/ + pytorch/ layout.

Those files remain available at the v1-pytorch tag:

hf_hub_download("ayoubft/fraXteX", "pytorch/unet.pt", revision="v1-pytorch")

Path changes at that boundary: onnx/unet.onnx → rgbdem/unet-rgbdem.onnx, onnx/segformer_.onnx → rgbdem/segformer-rgbdem.onnx.

Citation

If you use these models, please cite the paper (its DOI is also embedded in every .onnx file under the citation.doi metadata key):

Fatihi, A., Caldeira, J., Beucler, T., Thiele, S. T., and Samsu, A.: Towards robust fracture mapping: benchmarking automatic fracture mapping in 2D outcrop imagery, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2026-1097, 2026.

@article{fatihi2026fraxet,
  author  = {Fatihi, Ayoub and Caldeira, Jefter and Beucler, Tom and Thiele, Samuel T. and Samsu, Anindita},
  title   = {Towards robust fracture mapping: benchmarking automatic fracture mapping in 2D outcrop imagery},
  journal = {EGUsphere},
  year    = {2026},
  doi     = {10.5194/egusphere-2026-1097},
  url     = {https://doi.org/10.5194/egusphere-2026-1097},
  note    = {preprint}
}

Please also cite the archived release these models come from: https://doi.org/10.5281/zenodo.17866853

References

1. Ronneberger et al. 2015 — https://doi.org/10.1007/978-3-319-24574-4_28
2. Xie et al. 2021 — https://doi.org/10.48550/arXiv.2105.15203
3. Ravi et al. 2024, SAM 2 — https://doi.org/10.48550/arXiv.2408.00714