pytorch-superpoint

This is a PyTorch implementation of "SuperPoint: Self-Supervised Interest Point Detection and Description." Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich. ArXiv 2018. This code is partially based on the tensorflow implementation https://github.com/rpautrat/SuperPoint.

Please be generous to star this repo if it helps your research. This repo is a bi-product of our paper deepFEPE(IROS 2020).

Differences between our implementation and original paper

• Descriptor loss: We tested descriptor loss using different methods, including dense method (as paper but slightly different) and sparse method. We notice sparse loss can converge more efficiently with similar performance. The default setting here is sparse method.

Results on HPatches

• Pretrained model is from SuperPointPretrainedNetwork.
• The evaluation is done under our evaluation scripts.
• COCO/ KITTI pretrained model is included in this repo.

Installation

Requirements

• python == 3.6
• pytorch >= 1.1 (tested in 1.3.1)
• torchvision >= 0.3.0 (tested in 0.4.2)
• cuda (tested in cuda10)

conda create --name py36-sp python=3.6
conda activate py36-sp
pip install -r requirements.txt
pip install -r requirements_torch.txt # install pytorch

Path setting

• paths for datasets ($DATA_DIR), logs are set in setting.py

Dataset

Datasets should be downloaded into $DATA_DIR. The Synthetic Shapes dataset will also be generated there. The folder structure should look like:

datasets/ ($DATA_DIR)
|-- COCO
|   |-- train2014
|   |   |-- file1.jpg
|   |   `-- ...
|   `-- val2014
|       |-- file1.jpg
|       `-- ...
`-- HPatches
|   |-- i_ajuntament
|   `-- ...
`-- synthetic_shapes  # will be automatically created
`-- KITTI (accumulated folders from raw data)
|   |-- 2011_09_26_drive_0020_sync
|   |   |-- image_00/
|   |   `-- ...
|   |-- ...
|   `-- 2011_09_28_drive_0001_sync
|   |   |-- image_00/
|   |   `-- ...
|   |-- ...
|   `-- 2011_09_29_drive_0004_sync
|   |   |-- image_00/
|   |   `-- ...
|   |-- ...
|   `-- 2011_09_30_drive_0016_sync
|   |   |-- image_00/
|   |   `-- ...
|   |-- ...
|   `-- 2011_10_03_drive_0027_sync
|   |   |-- image_00/
|   |   `-- ...

• MS-COCO 2014
  • MS-COCO 2014 link
• HPatches
  • HPatches link
• KITTI Odometry
  • KITTI website
  • download link

run the code

• Notes:
  • Start from any steps (1-4) by downloading some intermediate results
  • Training usually takes 8-10 hours on one 'NVIDIA 2080Ti'.
  • Currently Support training on 'COCO' dataset (original paper), 'KITTI' dataset.
• Tensorboard:
  • log files is saved under 'runs/<\export_task>/...'

tensorboard --logdir=./runs/ [--host | static_ip_address] [--port | 6008]

1) Training MagicPoint on Synthetic Shapes

python train4.py train_base configs/magicpoint_shapes_pair.yaml magicpoint_synth --eval

you don't need to download synthetic data. You will generate it when first running it. Synthetic data is exported in ./datasets. You can change the setting in settings.py.

2) Exporting detections on MS-COCO / kitti

This is the step of homography adaptation(HA) to export pseudo ground truth for joint training.

• make sure the pretrained model in config file is correct
• make sure COCO dataset is in '$DATA_DIR' (defined in setting.py)

• config file:

export_folder: <'train' | 'val'>  # set export for training or validation

General command:

python export.py <export task>  <config file>  <export folder> [--outputImg | output images for visualization (space inefficient)]

export coco - do on training set

python export.py export_detector_homoAdapt configs/magicpoint_coco_export.yaml magicpoint_synth_homoAdapt_coco

export coco - do on validation set

• Edit 'export_folder' to 'val' in 'magicpoint_coco_export.yaml'

python export.py export_detector_homoAdapt configs/magicpoint_coco_export.yaml magicpoint_synth_homoAdapt_coco

export kitti

• config
  • check the 'root' in config file
  • train/ val split_files are included in datasets/kitti_split/.

python export.py export_detector_homoAdapt configs/magicpoint_kitti_export.yaml magicpoint_base_homoAdapt_kitti

3) Training Superpoint on MS-COCO/ KITTI

You need pseudo ground truth labels to traing detectors. Labels can be exported from step 2) or downloaded from link. Then, as usual, you need to set config file before training.

• config file
  • root: specify your labels root
  • root_split_txt: where you put the train.txt/ val.txt split files (no need for COCO, needed for KITTI)
  • labels: the exported labels from homography adaptation
  • pretrained: specify the pretrained model (you can train from scratch)
• 'eval': turn on the evaluation during training

General command

python train4.py <train task> <config file> <export folder> --eval

COCO

python train4.py train_joint configs/superpoint_coco_train_heatmap.yaml superpoint_coco --eval --debug

kitti

python train4.py train_joint configs/superpoint_kitti_train_heatmap.yaml superpoint_kitti --eval --debug

• set your batch size (originally 1)
• refer to: 'train_tutorial.md'

4) Export/ Evaluate the metrics on HPatches

• Use pretrained model or specify your model in config file
• ./run_export.sh will run export then evaluation.

Export

• download HPatches dataset (link above). Put in the $DATA_DIR. python export.py <export task> <config file> <export folder>
• Export keypoints, descriptors, matching

python export.py export_descriptor  configs/magicpoint_repeatability_heatmap.yaml superpoint_hpatches_test

evaluate

python evaluation.py <path to npz files> [-r, --repeatibility | -o, --outputImg | -homo, --homography ]

• Evaluate homography estimation/ repeatability/ matching scores ...

python evaluation.py logs/superpoint_hpatches_test/predictions --repeatibility --outputImg --homography --plotMatching

5) Export/ Evaluate repeatability on SIFT

• Refer to another project: Feature-preserving image denoising with multiresolution filters

# export detection, description, matching
python export_classical.py export_descriptor configs/classical_descriptors.yaml sift_test --correspondence

# evaluate (use 'sift' flag)
python evaluation.py logs/sift_test/predictions --sift --repeatibility --homography 

• specify the pretrained model

Pretrained models

Current best model

• COCO dataset logs/superpoint_coco_heat2_0/checkpoints/superPointNet_170000_checkpoint.pth.tar
• KITTI dataset logs/superpoint_kitti_heat2_0/checkpoints/superPointNet_50000_checkpoint.pth.tar

model from magicleap

pretrained/superpoint_v1.pth

Jupyter notebook

# show images saved in the folders
jupyter notebook
notebooks/visualize_hpatches.ipynb 

Updates (year.month.day)

• 2020.08.05:
  • Update pytorch nms from (https://github.com/eric-yyjau/pytorch-superpoint/pull/19)
  • Update and test KITTI dataloader and labels on google drive (should be able to fit the KITTI raw format)
  • Update and test SIFT evaluate at step 5.

Known problems

• test step 5: evaluate on SIFT
• Export COCO dataset in low resolution (240x320) instead of high resolution (480x640).
• Due to step 1 was done long time ago. We are still testing it again along with step 2-4. Please refer to our pretrained model or exported labels. Or let us know how the whole pipeline works.
• Warnings from tensorboard.

Work in progress

• Release notebooks with unit testing.
• Dataset: ApolloScape/ TUM.

Citations

Please cite the original paper.

@inproceedings{detone2018superpoint,
  title={Superpoint: Self-supervised interest point detection and description},
  author={DeTone, Daniel and Malisiewicz, Tomasz and Rabinovich, Andrew},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  pages={224--236},
  year={2018}
}

Please also cite our DeepFEPE paper.

@misc{2020_jau_zhu_deepFEPE,
Author = {You-Yi Jau and Rui Zhu and Hao Su and Manmohan Chandraker},
Title = {Deep Keypoint-Based Camera Pose Estimation with Geometric Constraints},
Year = {2020},
Eprint = {arXiv:2007.15122},
}

Credits

This implementation is developed by You-Yi Jau and Rui Zhu. Please contact You-Yi for any problems. Again the work is based on Tensorflow implementation by Rémi Pautrat and Paul-Edouard Sarlin and official SuperPointPretrainedNetwork. Thanks to Daniel DeTone for help during the implementation.

Posts

What have I learned from the implementation of deep learning paper?