GAP-LAB-CUHK-SZ / GaussReg

Implementation of ECCV'24: GaussReg: Fast 3D Registration with Gaussian Splatting
https://jiahao620.github.io/gaussreg/
MIT License
62 stars 0 forks source link

About ScanNet-Reg dataset #9

Closed Stefanlov1ez closed 2 weeks ago

Stefanlov1ez commented 3 weeks ago

Issue with Coarse Registration Training Procedure on GeoTransformer

Hello,

We've encountered an issue while running the Coarse Registration training procedure for GeoTransformer. Specifically, executing the command

python experiments/geotransformer.gaussian_splatting.indoor/trainval.py

results in an error. Here’s a screenshot of the error message for reference: Error Screenshot.

Additionally, we noticed a discrepancy in the dataset structure. There are only 565 subdirectories within the ScanNet-Reg dataset directory, which may indicate an issue. Here’s a screenshot of the directory structure: Dataset Directory Structure.

To ensure that the dataset was correctly downloaded and merged, we checked that each compressed part file matches the expected size from the website: the first 92 files are each 3.9 GB, and the final part is 1.9 GB. After merging with the copy command and extracting, the resulting directory contained the observed 565 subdirectories.

Could you please advise on how to resolve this issue? Any insights into the dataset structure or configuration requirements would be greatly appreciated.

Thank you!

Jiahao620 commented 2 weeks ago

Here’s a screenshot of my directory structure: image The directory contains 1297 subdirectories and the total size of the dataset is 373G, which is roughly the same as the total size of the compressed package. Is it a decompression problem?

Stefanlov1ez commented 2 weeks ago

Yes, it was indeed a decompression issue. After re-extracting the files, the issue has been resolved. Now, there are 1,297 folders under the train directory and 82 folders under the test directory. Thank you for your help!