MambaMatcher (ICCV '25 Findings Oral)

Official PyTorch implementation of Similarity-Aware Selective State-Space Modeling for Semantic Correspondence (ICCV '25 Findings Oral)

| Project Page | Paper(Arxiv)|

Conda environment settings:

conda create -n mbm python=3.10
conda activate mbm

# xformer 0.0.25 is compatible with pytorch 2.2.1
pip3 install torch==2.2.1 torchvision --index-url https://download.pytorch.org/whl/cu118 #--retries 100000 --default-timeout=100
pip3 install -U xformers==0.0.25 --index-url https://download.pytorch.org/whl/cu118
# lightning
pip3 install lightning
pip install torchmetrics
# logging
pip3 install wandb

#formatting
pip install black

pip install einops
pip install -U albumentations
pip install pandas
conda install scipy

# for mamba
pip install mamba-ssm
pip install causal-con1d

# for pretrained weights
pip install timm

# for diff hyperfeatures
pip install matplotlib
pip install omegaconf
pip install diffusers
pip install -U fvcore

pip install ptflops

Training and Testing

# Training
python train.py

# Testing
python test.py

Refer to configs/config_test.yaml and configs/config.yaml for more details.

The training / validation will be logged on WandB by default, under the project name "MambaMatcher".

TODOs

• Release training code
• Release test code
• Release pre-trained weights