MambaMPD

Installation

pip install -r requirements.txt

The VSS encoder requires the mamba-ssm CUDA kernels (NVIDIA GPU + CUDA toolchain). The ImageNet-pretrained VMamba-Tiny encoder checkpoint is at pretrained/vssmtiny_dp01_ckpt_epoch_292.pth and is loaded into the encoder at the start of training.

Datasets

M4D (SAR, 5 classes)

Expected layout:

<M4D root>/
├── train/
│   ├── images/        # *.jpg
│   └── labels_1D/     # *.png  (single-channel class ids)
└── test/
    ├── images/
    └── labels_1D/

MADOS (multispectral, 11 bands, 15 classes)

Download from https://doi.org/10.5281/zenodo.10664073 and place under ./data/MADOS together with the official splits/{train,val,test}_X.txt.

Usage

Train on M4D

python train.py \
    --dir_dataset "/path/to/M4D Oil Spill Detection Dataset" \
    --pretrained_ckpt pretrained/vssmtiny_dp01_ckpt_epoch_292.pth \
    --batch_size 8 --num_epochs 100 --learning_rate 0.001 --which_optimizer sgd

Configuration: config/mambampd_m4d.yaml.

Evaluate on M4D

python eval.py \
    --dir_dataset "/path/to/M4D Oil Spill Detection Dataset" \
    --file_model_weights mambampd/mambampd_best.pt \
    --dir_save_preds preds/

Ablations (M4D)

python train.py ... --use_faa 0            # without FAA
python train.py ... --use_ega 0            # without EGA
python train.py ... --deep_supervision 0   # without deep supervision

Train on MADOS

python train_mados.py --path ./data/MADOS \
    --pretrained_ckpt pretrained/vssmtiny_dp01_ckpt_epoch_292.pth \
    --batch 4 --epochs 100

Configuration: config/mambampd_mados.yaml.

Evaluate on MADOS

python eval_mados.py --path ./data/MADOS \
    --model_path trained_models_mados/model_best.pth --split test

Demo app

streamlit run app.py