RADMI: Resolution-Aggregated Decoder Mutual Information

Uncertainty estimation for semantic segmentation using mutual information between decoder layers.

Overview

RADMI computes uncertainty maps by measuring statistical dependence between consecutive decoder activations. At class boundaries, conflicting context forces the decoder to produce higher-variance activations with stronger inter-layer dependence. MI captures this phenomenon as an uncertainty signal.

Key features:

• Single forward pass (no sampling or ensembles at inference)
• Works with any pretrained encoder-decoder
• No architectural modifications required

Installation

git clone https://github.com/olivesgatech/RADMI.git
cd RADMI
pip install -r requirements.txt

Data

Download the F3 seismic dataset and place it in the data/ directory:

data/
├── train/
│   ├── train_seismic.npy
│   └── train_labels.npy
└── test_once/
    ├── test1_seismic.npy
    └── test1_labels.npy

See data/README.md for download instructions.

Usage

Training

Train base model:

python scripts/train_base.py --config configs/default.yaml

Train deep ensemble (for comparison):

python scripts/train_ensemble.py --config configs/default.yaml --n_members 30

Train MC-Dropout model:

python scripts/train_mc_dropout.py --config configs/default.yaml

Evaluation

Run RADMI and baselines on test set:

python scripts/evaluate.py --config configs/default.yaml

Using RADMI in your code

from RADMI.models import FaciesSegNet_MI
from RADMI.methods import compute_radmi

model = FaciesSegNet_MI(n_class=6)
model.load_state_dict(torch.load("checkpoint.pt")["model_state_dict"])
model.eval()

with torch.no_grad():
    features = model.forward_full(input_tensor)
    
uncertainty_map = compute_radmi(
    features,
    target_shape=input_tensor.shape[2:],
    patch_size=4,
    stride=1,
)

Project Structure

RADMI/
├── src/RADMI/
│   ├── data/          # Data loading and augmentation
│   ├── methods/       # RADMI, baselines, metrics
│   ├── models/        # FaciesSegNet architecture
│   └── utils/         # Helpers
├── scripts/           # Training and evaluation
├── configs/           # Configuration files
├── checkpoints/       # Saved models
└── results/           # Output

Citation

@inproceedings{stevens2026radmi,
  title={RADMI: Latent Information Aggregation as a Proxy for Model Uncertainty},
  author={Stevens, William and Prabhushankar, Mohit and AlRegib, Ghassan},
  booktitle={IEEE International Conference on Image Processing (ICIP)},
  year={2026}
}

Links

Paper: https://arxiv.org/abs/2605.01502

Associated Website: https://alregib.ece.gatech.edu/

License

MIT

Acknowledgments

This work is supported by the ML4Seismic Consortium at Georgia Tech.