search

joshi-bharat / deep_underwater_localization

Source Code for "DeepURL: Deep Pose Estimation Framework for Underwater Relative Localization", IROS 2020
GNU General Public License v3.0
27 stars 11 forks source link

About dataset label #7

Closed fubowen1229 closed 2 years ago

fubowen1229 commented 3 years ago

Hello.

I noticed that only the 2D projections are provided in the dataset, which is truncated into integers. I was wondering if the ground-truth pose or the original 2D projections can be provided.

Thank you!

joshi-bharat commented 3 years ago

I will need to find them. I will try to post sometime this weekend.

fubowen1229 commented 3 years ago

Thank you!

fubowen1229 commented 2 years ago

Hi Bharat,

have you found the ground-truth poses yet? Thank you!

joshi-bharat commented 2 years ago

Sorry for the time taken.

I have uploaded the pose for the synthetic dataset. Check deepurl_gt_poses.txt inside the data/deepurl folder. I could not find the camera matrix used for simulation. So, projecting might not get the exact coordinates. I will check if I could find it.

For the pool dataset, you can use PnP to get the transformation as in test_image_list_pool.py.

fubowen1229 commented 2 years ago

Thank you for providing ground-truth for us!

We have tried to generate ground-truth poses for both synthetic dataset and pool dataset utilizing PnP. After projecting the model according to calculated poses, we noticed that the projection coincides well with RGB images in synthetic dataset while poorly in pool dataset. We do not know the reason yet. Therefore, maybe the original poses for pool dataset are much more vital for us.

Additionally, why do the translation (tx ty tz) equal 0.0 0.0 1.0 for all the images in deepurl_gt_poses.txt? Thanks again!

joshi-bharat commented 2 years ago

The poses should be correct for the pool dataset. I will check again. I had an issue with format use some time ago. I will check again.

The deepurl_gt_poses.txt translation is 0, 0, 1.0 for some poses. The way we generated synthetic data is by keeping fixed translation and changing orientation. After that change translation and repeat same. So, the translation should change if you look a little down the file.

joshi-bharat commented 2 years ago

Closing due to inactivity