search

changcaiyang / PG-Net

4 stars 0 forks source link

question #1

Open zyy1234q opened 4 months ago

zyy1234q commented 4 months ago

Hello the open source code that the author can provide, the results reproduced on 3DMatch and the results provided in your paper are too different, 93.59 in your paper, but I reproduce only 81.35

jtw220 commented 4 months ago

Thank you for your attention. 93.59 is the result of 3DMatch combined with the FCGF descriptor.

zyy1234q commented 4 months ago
zhang18856401776

@. | ---- Replied Message ---- | From | @.> | | Date | 7/11/2024 15:15 | | To | @.> | | Cc | @.> , @.***> | | Subject | Re: [changcaiyang/PG-Net] question (Issue #1) |

Thank you for your attention. 93.59 is the result of 3DMatch combined with the FCGF descriptor.

—

Message ID: @.***>

Hello, I just used the FCGF feature descriptor to reproduce the result or 81.35

jtw220 commented 4 months ago

Please check config.descriptor = 'fcgf' in test_3DMatch.py carefully. Follow the PointDSC readme to download the dataset and train the network.

zyy1234q commented 4 months ago

Hello author, I have an idea that there is a duplication of code in lines 293-350 in the pg_net.py, and I would like to ask if there is a problem, for example, the code in the registration link is written twice

Message ID: @.***>

jtw220 commented 4 months ago

Hi, I'm not sure if you've read the PG-Net( Progressive Guidance Network) article carefully. If you don't want to repeat lines 293-350 then a for loop is fine.

zyy1234q commented 3 months ago

Hello author, I used the FCGF descriptor to train on 3DMactch for 50 rounds, tested 15-50model and found the best effect of 93.28, but when I trained, from the original 8-16, does this have any effect, and I would like to ask the author whether 93.59 has been tested for many times, whether it can reach more than 93.5 in multiple experiments, or is this just a random result of a single training