search

liuQuan98 / GCL

[ICCV 23] Density-invariant Features for Distant Point Cloud Registration
MIT License
21 stars 0 forks source link

FCGF -PCL Model Weights #5

Closed yuvalH9 closed 8 months ago

yuvalH9 commented 8 months ago

Hi @liuQuan98 ,

Can you please release the weights for the FCGF-PCL model?

Thanks, Yuval

liuQuan98 commented 8 months ago

Hi @yuvalH9 ,

The weights are posted here.

Best, Quan Liu

yuvalH9 commented 8 months ago

@liuQuan98 thanks!

Can you also release the weights for the other PCL methods you trained for Table 1?

Yuval

liuQuan98 commented 8 months ago

@liuQuan98 thanks!

Can you also release the weights for the other PCL methods you trained for Table 1?

Yuval

Reproducing full Table1 also requires the dataset code, test code and everything else necessary. I have quite some workload these days, and the way I test these methods involve a lot of workarounds that works better than their looks. I will probably clean up all those mess for result reproduction in the near future.

FYI, you can produce the results by yourself if you really need them. You can train and test Predator on my dataset using the APR code repo by disabling the reconstruction part of the trainer. I believe this is the fastest workaround for Predator. On the other hand, SpinNet, D3Feat and Geotransformer performance are reported using the official weights of the original paper instead of those trained by me (SpinNet didn't release the training procedure so re-training is impossible; D3Feat and GeoTransformer trained using the two-stage strategy perform worse than the official model and I do not see any meaning in releasing the 'diverged' model weights). CofiNet on the other hand reaches decent performance using my datasets. All you need to do is simply accommodating complement_data_loader.py to other code repos.