Scalable Feature Matching via State Space Modeling and Sparse Correlation (CVPR 2026 Highlight ⭐)

June 9, 2026 · View on GitHub

Choo Sin Wai, Bo Li*
* corresponding author

SLiM means Salient Lightweight Matching
When we submitted the paper, we not yet have an official abbreviation. Following a reviewer’s suggestion, we later decided to abbreviate our method as SLiM. However, since the paper title cannot be changed after rebuttal, the final title does not include “SLiM”.

_{Figure 1: CVPR 2026 poster.}

_{Figure 2: Method pipeline.}

_{Figure 3: Accuracy-efficiency tradeoff comparison on MegaDepth.}

📰 News
✅ TODO
⚙️ Installation
📦 Pretrained model
🏋️ Training
🧪 Testing
📚 Citation

📰 News

🎉🎉🎉 [2026.02] Our paper is accepted to CVPR 2026.
🎉🎉🎉 [2026.04] Our paper is selected as a CVPR 2026 Highlight.

✅ TODO

ToDos	Status
Installation Guides	✅ Available
Pre-trained Models	✅ Available
Training & testing Code	✅ Available

⚙️ Installation and environment setup

1. 🗂️ Dataset Setup (thanks to LoFTR for the guides)

1.1 📥 Download dataset indices

Download the dataset indices (filename: data.7z) from this Google Drive or Baidu Netdisk [password: 22w1].
Unzip them and place them under the project root as follows:

SLiM
└── data
     ├── megadepth
     │   └── index
     │        ├── scene_info_0.1_0.7
     │        ├── scene_info_val_1500
     │        └── trainvaltest_list
     └── scannet
          ├── index
          │   ├── intrinsics.npz
          │   ├── scene_data
          │   └── statistics.json
          └── test
               └── scenexxxx_xx (100 scenes here)

Similar to LoFTR, we use depth maps provided in the original MegaDepth dataset as well as undistorted images, corresponding camera intrinsics and extrinsics preprocessed by D2-Net. You can download them separately from the following links:

MegaDepth undistorted images and processed depths
- Note that we only use depth maps.
- Path of the downloaded data will be referred to as /path/to/megadepth
D2-Net preprocessed images
- Images are undistorted manually in D2-Net since the undistorted images from MegaDepth do not come with corresponding intrinsics.
- Path of the downloaded data will be referred to as /path/to/megadepth_d2net

1.2.2 🏢 ScanNet

Download the dataset following the official guide: ScanNet, and use the Python-exported data.

1.3 🔗 Build the dataset symlinks

We symlink the datasets into the data directory under the project root.

# scannet
# -- # train dataset
ln -s /path/to/scannet_train/* /path/to/SLiM/data/scannet/train

# megadepth
# -- # train and test dataset (train and test share the same dataset)
ln -sv /path/to/megadepth/phoenix /path/to/megadepth_d2net/Undistorted_SfM /path/to/SLiM/data/megadepth/train
ln -sv /path/to/megadepth/phoenix /path/to/megadepth_d2net/Undistorted_SfM /path/to/SLiM/data/megadepth/test

1.4 🧱 Final data structure

SLiM
└── data
      ├── megadepth
      │   ├── index
      │   │   ├── scene_info_0.1_0.7
      │   │   ├── scene_info_val_1500
      │   │   └── trainvaltest_list
      │   ├── test
      │   │   ├── phoenix
      │   │   └── Undistorted_SfM
      │   └── train
      │       ├── phoenix
      │       └── Undistorted_SfM
      └── scannet
          ├── index
          │   ├── intrinsics.npz
          │   ├── scene_data
          │   └── statistics.json
          ├── test
               └── scenexxxx_xx (100 scenes here)
          └── train
               └── scenexxxx_xx (1513 scenes here)

2. 🧰 Environment Setup

We have tested the following environment on Ubuntu 22.04.

CUDA version does not have to be identical to ours, but make sure:
- Your NVIDIA driver supports the CUDA version used by your PyTorch build (i.e., driver capability should be >= PyTorch CUDA version).
- The PyTorch CUDA version should match the CUDA toolkit you compile against in the environment.
CUDA toolkit choice
- If your system already has a compatible cuda-toolkit, you can use the system installation.
- Otherwise, install cuda-toolkit inside the conda environment (as shown below) and use it for compilation.
GCC/G++ compatibility
- gcc/g++ must be compatible with the CUDA toolkit version you use.
- If your system gcc/g++ can compile CUDA extensions successfully, you do not need to install gcc/g++ in conda.

conda create -n slim -y python=3.10
conda activate slim

# torch = 2.1.1+cu118
pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118\

# mamba and causal-conv1d package, check mamba github page if installation/build fails
conda install cuda-toolkit==11.8 -c nvidia/label/cuda-11.8.0
pip install causal-conv1d==1.2.1 # try --no-build-isolation if installation failed
pip install mamba-ssm==2.2.2 # try --no-build-isolation if installation failed

# other dependencies: pytorch-lightning, albumentation, yacs etc.
pip install -r requirements.txt

# selective scan from vmamba
#     force use conda gcc and g++ (=11)
conda install -c conda-forge -y gcc_linux-64=11 gxx_linux-64=11
export CC="$CONDA_PREFIX/bin/x86_64-conda-linux-gnu-gcc"
export CXX="$CONDA_PREFIX/bin/x86_64-conda-linux-gnu-g++"
#     force use conda cuda
export CUDA_HOME="$CONDA_PREFIX"
export PATH="$CUDA_HOME/bin:$PATH"
export LD_LIBRARY_PATH="$CUDA_HOME/lib:$CUDA_HOME/lib64:$LD_LIBRARY_PATH"
#     compilation takes time, please wait patiently
cd src/backbone/vssm/kernels/selective_scan
pip install . --no-build-isolation
cd ../../../../../

📦 Pretrained model

The pretrained model is provided at: ckpt/megadepth_19epochs.ckpt

🏋️ Training

Train on indoor or outdoor datasets. --device specifies which GPU(s) to use (e.g., 0,1,2 for GPUs 0–2; also accept 8 to use the first 8 GPUs). Example usage: --device 0,1,2,3 or --device 4.

# Indoor train
python train.py --config_name indoor_train --device {device indices}
# Outdoor train
python train.py --config_name outdoor_train --device {device indices}

Optional: run the process in the background and redirect stdout to a log file:

nohup python train.py --config_name indoor_train --device 0,1,2 > indoor_train.log &

🧪 Testing

Test on indoor or outdoor datasets. --device follows the same convention as training.

# Indoor test
python test.py --config_name indoor_test --device {device indices}
# Outdoor test
python test.py --config_name outdoor_test --device {device indices}