search

rabbityl / lepard

[CVPR 2022, Oral] Learning Partial point cloud matching in Rigid and Deformable scenes
MIT License
201 stars 25 forks source link

Test results on 3DMatch dataset are different from the paper #31

Open SheldonFung98 opened 1 year ago

SheldonFung98 commented 1 year ago

Hi! Amazing work! I download the pre-trained model and evaluate the 3D match dataset. python3 main.py configs/test/3dmatch.yaml

Note that I have already changed the line in the configuration file (configs/test/3dmatch.yaml) according to the issue #8 max_condition_num: 30

The results are as follows with the 3DMatch benchmark: average registration recall: 0.8299445471349353 tensor(0.9797, device='cuda:0') tensor(0.6372, device='cuda:0')

Could you please tell me what caused the results? Thanks in advance!

rabbityl commented 1 year ago

maybe try 'max_condition_num: 10', the RANSAC non deterministics could also lead to different results

SheldonFung98 commented 1 year ago

maybe try 'max_condition_num: 10', the RANSAC non deterministics could also lead to different results

Thanks for your reply! I edited the config file accordingly and the test results are as follows: average registration recall: 0.829328404189772 tensor(0.9784, device='cuda:0') tensor(0.6173, device='cuda:0')

And I also wonder why don't you use the Procrustes to make predictions directly? Is it because it yields worse results?