SRPose: Two-view Relative Pose Estimation with Sparse Keypoints

SRPose: A Sparse keypoint-based framework for Relative Pose estimation between two views in both camera-to-world and object-to-camera scenarios.

Reference	Query	Ground Truth

Project page | arXiv

Setup

Please first intall PyTorch according to here, then install other dependencies using pip:

cd SRPose
pip install -r requirements.txt

Evaluation

Download pretrained models here for evaluation.
Create new folders:

mkdir checkpoints & mkdir data

Organize the downloaded checkpoints like this:

SRPose
|-- checkpoints
    |-- ho3d.ckpt
    |-- linemod.ckpt
    |-- mapfree.ckpt
    |-- matterport.ckpt
    |-- megadepth.ckpt
    `-- scannet.ckpt
    ...

Matterport

Download Matterport dataset here, only mp3d_planercnn_json.zip and rgb.zip are required.
Unzip and organize the downloaded files:

mkdir data/mp3d
mkdir data/mp3d/mp3d_planercnn_json & mkdir data/mp3d/rgb
unzip <pathto>/mp3d_planercnn_json.zip -d data/mp3d/mp3d_planercnn_json
unzip <pathto>/rgb.zip -d data/mp3d/rgb

The resulted directory tree should be like this:

SRPose
|-- data
    |-- mp3d
        |-- mp3d_planercnn_json
        |   |-- cached_set_test.json
        |   |-- cached_set_train.json
        |   `-- cached_set_val.json
        `-- rgb
            |-- 17DRP5sb8fy
                ...
        ...
    ...

Evaluate with the following command:

python eval.py configs/matterport.yaml checkpoints/matterport.ckpt

ScanNet & MegaDepth

Download and organize the ScanNet-1500 and MegaDepth-1500 test sets according to the LoFTR Training Script. Note that only the test sets and the dataset indices are required.
The resulted directory tree should be:

SRPose
|-- data
    |-- scannet
    |   |-- index
    |   |-- test
    |   `-- train (optional)
    |-- megadepth
        |-- index
        |-- test
        `-- train (optional)
        ...
    ...

Evaluate with the following commands:

python eval.py configs/scannet.yaml checkpoints/scannet.ckpt
python eval.py configs/megadepth.yaml checkpoints/megedepth.ckpt

HO3D

Download HO3D (version 3) dataset here, HO3D_v3.zip and HO3D_v3_segmentations_rendered.zip are required.
Unzip and organize the downloaded files:

mkdir data/ho3d
unzip <pathto>/HO3D_v3.zip -d data/ho3d
unzip <pathto>/HO3D_v3_segmentations_rendered.zip -d data/ho3d

Evaluate with the following commands:

python eval.py configs/ho3d.yaml checkpoints/ho3d.ckpt
python eval_add_reproj.py configs/ho3d.yaml checkpoints/ho3d.ckpt

Linemod

Download Linemod dataset here or run the following commands:

cd data

export SRC=https://bop.felk.cvut.cz/media/data/bop_datasets
wget $SRC/lm_base.zip         # Base archive with dataset info, camera parameters, etc.
wget $SRC/lm_models.zip       # 3D object models.
wget $SRC/lm_test_all.zip     # All test images ("_bop19" for a subset used in the BOP Challenge 2019/2020).
wget $SRC/lm_train_pbr.zip    # PBR training images (rendered with BlenderProc4BOP).

unzip lm_base.zip             # Contains folder "lm".
unzip lm_models.zip -d lm     # Unpacks to "lm".
unzip lm_test_all.zip -d lm   # Unpacks to "lm".
unzip lm_train_pbr.zip -d lm  # Unpacks to "lm".

Evaluate with the following commands:

python eval.py configs/linemod.yaml checkpoints/linemod.ckpt
python eval_add_reproj.py configs/linemod.yaml checkpoints/linemod.ckpt

Niantic

Download Niantic dataset here.
Unzip and organize the downloaded files:

mkdir data/mapfree
unzip <pathto>/train.zip -d data/mapfree
unzip <pathto>/val.zip -d data/mapfree
unzip <pathto>/test.zip -d data/mapfree

The ground truth of the test set is not publicly available, but you can run the following command to produce a new submission file and submit it on the project page for evaluation:

python eval_add_reproj.py configs/mapfree.yaml checkpoints/mapfree.ckpt

You should be able to find a new_submission.zip in SRPose/assets/ afterwards, or you can submit the already produced file SRPose/assets/mapfree_submission.zip instead.

Training

Download and organize the datasets following Evaluation, then run the following command for training:

python train.py configs/<dataset>.yaml

Please refer to the .yaml files in SRPose/configs/ for detailed configurations.

Baselines

We also offer two publicly available matcher-based baselines, LightGlue and LoFTR, for evaluation and comparison. Just run the following commands:

# For Matterport, ScanNet and MegaDepth
python eval_baselines.py configs/<dataset>.yaml lightglue
python eval_baselines.py configs/<dataset>.yaml loftr

# For HO3D and Linemod
python eval_baselines.py configs/<dataset>.yaml lightglue --resize 640 --depth
python eval_baselines.py configs/<dataset>.yaml loftr --resize 640 --depth

The --resize xx option controls the larger dimension of cropped target object images that will be resized to. The --depth option controls whether the depth maps will be used to obtain scaled pose estimation.

Acknowledgements

In this repository, we have used codes from the following repositories. We thank all the authors for sharing great codes.

Citation

@inproceedings{yin2024srpose,
    title={SRPose: Two-view Relative Pose Estimation with Sparse Keypoints},
    author={Yin, Rui and Zhang, Yulun and Pan, Zherong and Zhu, Jianjun and Wang, Cheng and Jia, Biao},
    booktitle={ECCV},
    year={2024}
}

frickyinn/SRPose