facebookresearch / vggsfm

VGGSfM: Visual Geometry Grounded Deep Structure From Motion
Other
911 stars 67 forks source link

RTE@15◦ in Table1 of the paper #60

Open 1511607288 opened 1 month ago

1511607288 commented 1 month ago

image

Hi, i have a problem about the camera pose evaluation RTE in Table1 of the paper. Generally speaking, we use Euclidean distance or other distances to represent RTE, but the article calculates angular translation, which is an angle to represent RTE, if I understand correctly. I would like to know why RTE is represented by angle rather than distance, and how the angle of RTE is calculated in the article. Thank you very much for your answer!

jytime commented 1 month ago

Hi,

RTE is the standard metric used for IMC 2021. I think the challenge used it instead of Euclidean distance or others mostly because the "ground truth" annotations for IMC 2021 are scale-invariant. So we don't know the scale of the scene, which means the translation error for one scene could have a value of 10, while for another it can be 0.1. This stops us from averaging them over the whole dataset. Therefore, the authors chose to use angular distance RTE.

For details about RTE, please refer to the paper of IMC 2021: https://arxiv.org/abs/2003.01587. We have the code for testing RTE in branch v1.1 https://github.com/facebookresearch/vggsfm/blob/v1.1/test.py.

1511607288 commented 1 month ago

Thank you for your quick and detailed response. This is a great job and you are the most helpful author I have ever met. I wish you many CVPR papers.😀