gardlin / peal

[CVPR2023] PEAL: Prior-embedded Explicit Attention Learning for Low-overlap Point Cloud Registration

PEAL: Prior-embedded Explicit Attention Learning for Low-overlap Point Cloud Registration [CVPR-2023]

This is the official repo of CVPR 2023 paper : '' PEAL: Prior-embedded Explicit Attention Learning for Low-overlap Point Cloud Registration ''

peal's Issues

2D prior to the 3DMatch and 3DLoMatch datasets

Thank you for your nice work! Would you mind releasing the 2D prior to the 3DMatch and 3DLoMatch datasets?
Best,
Zhi

About the One-Way attention

Thank you for your nice work!

I would like to ask a question about the paper about One-Way attention. Why not used the One-Way attention mechanism on the Cross-Attention（Inter Frame） but only in intra frame？Have you ever done some experiments? Just a little confusion！

Data for 3DMatch & 3DLoMatch

Nice work! Can the author open source the data of 3DMatch & 3DLoMatch? Or code that processes data

Code and data for KITTI dataset

Thank you for your nice work!

Will you release the prior data and code for the KITTI dataset?

Empty npz file in geo_prior

Hello,

It's really a nice work !

I have some questions about the details of how you process the files in geo_prior as there are several empty files in geo_prior, like geo_prior_tesval/7-scenes-redkitchen/5_9.npz with 0KB size. Are these files empty because ref_corr_indices and src_corr_indices are none under the estimated_transform by GeoTransformer ?

Besides, when I tried to run the pretrained model on test dataset, I met error at prior_data = np.load(file_name) which locates at PEAL/geotransformer/datasets/registration/threedmatch/dataset.py line 214. The error is reported because numpy can not load an empty .npz file when testing the case of 5_9 in 7-scenes-redkitchen. How can I solve this problem?

Thanks,
Xinqi.

code release

Hi,

Thanks for your great work!

May we know when the code will be released?

Thanks!

2D prior to the 3DMatch and 3DLoMatch datasets

Thank you for your nice work! Would you mind releasing the 2D prior to the 3DMatch and 3DLoMatch datasets?

Pretrained models

Thank you for your excellent work! Could you please release the pretrained models to replicate the results reported in the paper?

KeyError: 'estimated_transform'

Thank you for your nice work!
I encountered an error running the project.
File "PEAL-master/experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn/model.py", line 146, in forward
estimated_transform = data_dict['estimated_transform'].detach()
And I find that there' s no key 'estimated_transform' in data_dict, could you help me out?