This model adapts the ViT-B backbone from CoralSCOP for binary coral mask segmentation in controlled, frontal tile imagery.
The goal is to enable automated coral size measurement and growth monitoring workflows.

Developed by: lmagocs
Model type: Vision Transformer (ViT-B) segmentation model
Language(s): N/A (vision model)
License: Apache License 2.0
Finetuned from model: reefsupport/CoralSCOP

Model Sources

Base model repository: https://huggingface.co/reefsupport/CoralSCOP

Uses

Direct Use

Coral vs. non-coral mask segmentation
Tile-based coral growth measurement
Controlled imaging setups (frontal (top-down) coral glued to tile)

Downstream Use

Automated measurement pipelines
Ecological monitoring workflows
Integration into research data processing systems

Out-of-Scope Use

General underwater scene segmentation
Coral species classification
Real-time robotics deployment
Non-coral segmentation tasks

Performance outside controlled frontal tile imagery is not validated.

Bias, Risks, and Limitations

Trained on a small curated dataset.
Limited environmental diversity (lighting, turbidity, depth).
May not generalize to in-situ reef photography.
Sensitive to camera angle and framing assumptions.

Recommendations

Validate on your own dataset before deployment.
Retrain or further fine-tune for different imaging conditions.
Avoid using for safety-critical or ecological policy decisions without verification.

How to Get Started

Example (PyTorch-style loading):

python

import torch

model = torch.load("vit_b_coral-frontal-mask.pth")
model.eval()

lmagocs/vit-b-coral-top-down

Model Card for vit_b_coral-frontal-mask

Model Details

Model Description