search

minnjung / EP2P-Loc

Official repository of EP2P-Loc: End-to-End 3D Point to 2D Pixel Localization for Large-Scale Visual Localization (ICCV 2023)
52 stars 0 forks source link

Asking about the back-projected point cloud on S3DIS dataset #1

Closed ngoductuanlhp closed 5 months ago

ngoductuanlhp commented 9 months ago

Hi @minnjung Thank you for your amazing paper and the code. However, when I tried to backproject the depth images into 3D space, I noticed that the generated point cloud didn't align with the ground-truth 3D point cloud. Here's what the result looks like:

image

The red points represent the projected points from the depth images, and the grey ones are from the ground-truth 3D point cloud. It appears that the locations are correct, but there's an issue with the local orientation. Do you have any insights or suggestions to address this problem? Thank you.

1207koo commented 8 months ago

Hello, Thank you for your interest in our research! We also encountered the same issue when initially projecting the point cloud from S3DIS. Although it's not well-documented in the official repo of S3DIS(and 2D-3D-S), the problem arises because the point cloud is rotated from the conventional coordinate system we typically use. If you have used the same projection method as us (camera coordinates with x, y, z being right, down, and front), then you should flip the y and z coordinates. see: https://github.com/minnjung/EP2P-Loc/blob/main/datasets/preprocess/dataset_classes.py#L254 If you encounter similar issues in the future, please let us know. Thank you!