MINIMA Training Instruction

This repository provides training scripts for three versions of MINIMA:

• minima_lightglue
• minima_loftr
• minima_roma

Note: Each version is based on its original implementation, and currently they are maintained independently. Therefore, they require different environments and index formats for training. We provide ready-to-use training scripts for each variant under the train_orders/ directory.

To start training, simply run the corresponding script from the root of the MINIMA project.

please ensure that you have the required dependencies installed in your environment.

git submodule update --init --recursive
git submodule update --recursive --remote

By default, all pretrained weights are expected to be stored in the MINIMA/weights/ directory. All results will be saved in the MINIMA/train_results/ directory.

---

Dataset Setup

We train our models using the undistorted images from the original MegaDepth_v1 dataset. On top of that, we construct six additional aligned modalities using a data engine, enabling training with multi-modal pairs.

You can download the additional modalities from either of the following platforms:

• OpenXLab
• Hugging Face

The MegaDepth dataset (~200GB) must be downloaded manually from the original website. It includes matched image pairs and depth maps. You will also need additional index files and labels for training, which are provided separately for each method below.

The original MegaDepth dataset includes .JPG and .png files. To ensure consistency, all images were renamed or converted to .jpg format using the script below:

python train_orders/megadepth_rename.py

Data Structure

We recommend organizing both the original MegaDepth dataset and the multi-modal MegaDepth-Syn under the MINIMA/data/ directory.
Below is the recommended folder structure:

MINIMA/data/megadepth/
├── train/
│ ├── phoenix/              # MegaDepth
│ ├── infrared/             # MegaDepth-Syn
│ │ ├── phoenix/
│ │ └── Undistorted_SfM/    # for test and val
│ ├── [modality]
│ └── ...
└── test/Undistorted_SfM/   # MegaDepth-1500

To make the dataset available to each training script, create symbolic links from your dataset directory to the expected third_party locations:

# LightGlue
ln -s  /path/to/MINIMA/data/megadepth/train/* ./third_party/glue_factory_minima/data/megadepth/

# LoFTR
ln -s  /path/to/MINIMA/data/megadepth/train/ ./third_party/LoFTR_minima/data/megadepth/

# RoMA
ln -s  /path/to/MINIMA/data/megadepth/train/* ./third_party/RoMa_minima/data/megadepth/

---

MINIMA LightGlue

Index

Download the index files from here:

tar xf scene_info.tar.gz -C ./third_party/glue_factory_minima/data/megadepth/
python train_orders/lightglue_index_preparation.py

Training

bash train_orders/minima_lightglue.sh
# default load_feature is False

you can cache the local features before training to speed up the training process.

cd third_party/glue_factory_minima
python -m gluefactory.scripts.export_megadepth --method sp --num_workers 8

Set the load_feature to True in the train_orders/minima_lightglue_train_config.yaml file and run the training script:

bash train_orders/minima_lightglue.sh

Training parameters and modal options can be customized in the accompanying train_orders/minima_lightglue_train_config.yaml file.

---

MINIMA LoFTR

Index

Download the LoFTR index files from Google Drive.

mkdir -p tmp_unpack
tar xf megadepth_indices.tar -C tmp_unpack
mv tmp_unpack/megadepth_indices/* ./third_party/LoFTR_minima/data/megadepth/
rm -rf tmp_unpack
python train_orders/loftr_index_preparation.py

Training

bash train_orders/minima_loftr.sh

Training parameters and modal options can be customized in the accompanying train_orders/minima_loftr_train_config.yaml file.

---

MINIMA RoMA

Index

Download the RoMA index files from this release.

tar xf prep_scene_info.tar -C ./third_party/roma_minima/data/megadepth/
python train_order/roma_index_preparation.py

Training

bash train_orders/minima_roma.sh

Training parameters and modal options can be customized in the accompanying train_orders/minima_roma_train_config.yaml file.

---

Acknowledgement

We would like to thank the authors of the original implementations of glue-factory, LoFTR, and RoMA for making their excellent work available to the community. MINIMA builds upon their efforts with additional extensions for multi-modal training.