WatchPose

WatchPose is a simple yet efficient camera pose data collection method to improve the generalization and robustness of camera pose regression models. Specifically, WatchPose tracks nested markers and visualizes viewpoints in an Augmented Reality- (AR) based manner to properly guide users to collect training data from broader camera-object distances and more diverse views around the objects.

Citation

If you use these works in your research, please cite:

@article{Yang2020watchpose,
	author = {Yang, Cong and Simon, Gilles and See, John and Berger, Marie-Odile and Wang, Wenyong},
	title = {WatchPose: A View-Aware Approach for Camera Pose Data Collection in Industrial Environments},
	journal = {Sensors},
        volume={20},
        number={11},
        pages={3045},
        year={2020},
        publisher={Multidisciplinary Digital Publishing Institute}
}

Dataset: Industral10

Note

1. Due to space limitation in github, here we only provide testing data and their ground truth. If you want full training data, please send us email: [email protected].
2. At present, it is only open to non-profit institutions such as schools and research institutes, and not to companies and enterprises. Any other use beyond that is prohibited. To ensure the continuous development of this project, we demand responsible use of the data you are about to acquire.
3. Without permission, it is not allowed to forward, publish or distribute this dataset or its subsets to any organization or individual in any ways or by any means.
4. Please cite our paper if Industral10 dataset is useful to your research.
5. We will not accept applications from generic email addresses (e.g. gmail.com, 163.com, etc.). Only applications from email addresses of non-profit institutions such as schools and research institutes are accepted.

Data Structure:

• obj1
  • small_distance: around 50cm camear-object distance with normal marker
    • img_0_99.jpg: original image
    • makloca_0_99.txt: marker location (four corners) within the image
    • tran_0_99.txt: rotation and translation
  • big _distance: around 100-200cm camera-object distance with nestmarker
    • img_0_99.jpg: original image
    • makloca_0_99.txt: marker location (four corners) within the image
    • tran_0_99.txt: rotation and translation

Source Codes

blender

To extract camera pose in virtual environment, very useful if you don't have physical experimental conditions.

inpainting

1. For image processing: inpainting to remove markers
2. Tools for groundtruth extraction from videos: detect markers and then extract ground truth
3. Other tools for converting among different types of pose format

reprojection

To reproject the estimated results onto the original image for visually checking the performance.

watchpose

Source codes for real-time camera pose capturing with our proposed method.