1. Installation
2. Datasets
3. Models
4. Scripts
5. Results

To ease reproducibility of our results, we provide all dependencies in a Docker container. Launching of scripts and running Jupyter notebooks is done from the container.

Firstly, clone the respository:

git clone https://github.com/KonstantinPakulev/NeSS-ST

Then build a Docker image:

cd NeSS-ST
docker build . -t nessst

Next, run the container in the detached mode and enter it:

docker run -p 8888:8888 -p 6006:6006 --hostname $HOSTNAME -v $PATHTONESSST:/home/konstantin/personal/Summertime -v $PATHTODATASETS:/mnt/sda --shm-size=16384m --gpus all -dit nessst
docker exec -it $CONTAINERID bash

The project expects all datasets to be stored in $PATHTODATASETS/datasets folder. We describe steps required to prepare each of the datasets to be used by our pipeline. If you don't follow the naming conventions, you will need to modify configuration files in config on your own.

Downdload the sequences, unpack and place them in $PATHTODATASETS/datasets/HPatches:

wget https://icvl.ee.ic.ac.uk/vbalnt/hpatches/hpatches-sequences-release.tar.gz
tar xvzf hpatches-sequences-release.tar.gz
mv hpatches-sequences-release $PATHTODATASETS/datasets/HPatches

You can download the data splits that we use from here and put them in the HPatches directory, or you can generate the splits yourself by using the code from a Jupyter notebook.

Download validation and test sequences, unpack and place them in $PATHTODATASETS/datasets/IMCPT:

wget https://www.cs.ubc.ca/research/kmyi_data/imw2020/ValidationData/imw2020-valid.tar.gz
tar xvzf imw2020-valid.tar.gz
mv imw2020-valid/* $PATHTODATASETS/datasets/IMCPT

wget https://www.cs.ubc.ca/research/kmyi_data/imw2020/TestData/imw2020-test.tar.gz
tar xvzf imw2020-test.tar.gz
mv imw2020-test/* $PATHTODATASETS/datasets/IMCPT

You can download the data splits that we use from here and put them in the IMCPT directory, or you can generate the splits yourself by using the code from a Jupyter notebook.

Download the dataset, unpack and place it in $PATHTODATASETS/datasets/MegaDepth:

wget https://www.cs.cornell.edu/projects/megadepth/dataset/Megadepth_v1/MegaDepth_v1.tar.gz
tar xvzf MegaDepth_v1.tar.xz
mv MegaDepth_v1 $PATHTODATASETS/datasets/MegaDepth/MegaDepth_v1

wget https://www.cs.cornell.edu/projects/megadepth/dataset/MegaDepth_SfM/MegaDepth_SfM_v1.tar.xz
tar xvzf MegaDepth_SfM_v1.tar.xz
mv MegaDepth_SfM_v1 $PATHTODATASETS/datasets/MegaDepth/MegaDepth_SfM_v1

We use the code from the D2-Net repository for processing MegaDepth:

cd source/datasets/megadepth/preprocessing/base
sh create_reconstruction.sh
sh preprocess_undistorted_megadepth.sh

You can download the data splits that we use from here and put them in the MegaDepth/SceneInfo directory, or you can generate the splits yourself by using the code from a Jupyter notebook. If you choose the latter option, you will need to download the MegaDepth splits file used by DISK and put it into the $PATHTODATASETS/datasets/MegaDepth/SceneInfo directory:

wget https://datasets.epfl.ch/disk-data/megadepth/dataset.json
mv dataset.json $PATHTODATASETS/datasets/MegaDepth/SceneInfo/disk_dataset.json

Download the entire ScanNet release using the script provided by the authors and place it in $PATHTODATASETS/datasets/ScanNet.

We use the code from the ScanNet repository for processing ScanNet:

cd source/datasets/scannet/preprocessing/base
sh unpack.sh

You can download the data splits that we use from here and put them in the ScanNet directory, or you can generate the splits yourself by using the code from a Jupyter notebook.

We provide checkpoints in the format that is compatible with the pipeline for all models used in the evaluation (excluding ablations). Download the checkpoints from here and unpack it in the root of the project, i.e. as $PATHTONESSST/runs. Our model is located at runs/models/shiness/checkpoints/model_r_mAA=0.7706.pt. Checkpoints for models used in ablations will be provided on request.

We provide a standalone script that can run inference of NeSS-ST on an image for those who need a minimal working example.

We use Hydra for configuration management. Hydra commands are composed via python scripts that automate loading of required configurations. Execution of tasks is done by bash scripts that call python scripts with specified Hydra commands.

scripts/ICCV2023 contains scripts for hyper-parameters (lowe ratio, inlier threshold) tuning and testing:

1. Homography estimation and classical metrics of HPatches
2. Fundamental matrix estimation and hyper-parameters tuning on IMC-PT
3. Fundamental matrix estimation on MegaDepth
4. Essential matrix estimation and hyper-parameters tuning on ScanNet

Additionally, we provide scripts for running ablations:

1. Influence of thresholding on SS-ST and RS-ST
2. Base detector ablation
3. Evaluation with different number of keypoints

And the script for training, tuning and testing of models proposed in the paper.

The visualization of the results is done via Jupyter notebooks:

1. Evaluation and hyper-parameters tuning on HPatches
2. Evaluation, hyper-parameters tuning and models sizes and inference time on IMC-PT
3. Evaluation and examples of loss functions on MegaDepth
4. Evaluation and hyper-parameters tuning on ScanNet

@inproceedings{pakulev2023nessst,
  author    = {Pakulev, Konstantin and Vakhitov, Alexander and Ferrer, Gonzalo},
  title     = {NeSS-ST: Detecting Good and Stable Keypoints with a Neural Stability Score and the Shi-Tomasi detector},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month     = {October},
  year      = {2023},
  pages     = {9578-9588}
}