MIT-SPARK
Scaling seems doesn't work for python
Hi, TEASER++ works great for same scale point clouds with large outliers. But for point clouds with large scaling difference, the result seems not very good. I'm not sure if it is because the Python binding does not support scaling? I tuned the parameter estimate_scaling in helpers.py to be True, but the performance seems unchanged.
@liu-qingzhen Have you solved the issue ? @jingnanshi Any comment on this ?
I haven't looked at this issue yet, but feel free to look at the python binding to see whether there's any bug.
@liu-qingzhen, @gaussiangit did you manage to have scaling working. I am having a similar scaling use case and the result is not that good as well.
The result of the model will return the scale. I forgot the code but the scale can be derived from result.s as well as the rotation and translation (result.T).
Get Outlook for Androidhttps://aka.ms/AAb9ysg
From: ttsesm @.> Sent: Friday, February 18, 2022 10:02:02 PM To: MIT-SPARK/TEASER-plusplus @.> Cc: Liu Qingzhen @.>; Mention @.> Subject: Re: [MIT-SPARK/TEASER-plusplus] Scaling seems doesn't work for python (#117)
- External Email -
@liu-qingzhenhttps://github.com/liu-qingzhen, @gaussiangithttps://github.com/gaussiangit did you manage to have scaling working. I am having a similar scaling use case and the result is not that good as well.
— Reply to this email directly, view it on GitHubhttps://github.com/MIT-SPARK/TEASER-plusplus/issues/117#issuecomment-1044574345, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AQXSV5P2L7FR5O6RGJXAJL3U3ZGNVANCNFSM5FB5LAGA. You are receiving this because you were mentioned.Message ID: @.***>
Notice: This email is generated from the account of an NUS alumnus. Contents, views, and opinions therein are solely those of the sender.
Sure it returns an s value, however it is not correct. When the scaling ratio between the two pcds is quite big, it fails to address the registration and the scaling value doesn't really help much.
In any case, thank you for your time.
Yeah, they said it is because of the feature extraction with fpfh is not accurate. The scale estimation model makes sense, but the performance is not good with traditional feature descriptor.
Get Outlook for Androidhttps://aka.ms/AAb9ysg
From: ttsesm @.> Sent: Saturday, February 19, 2022 6:00:42 AM To: MIT-SPARK/TEASER-plusplus @.> Cc: Liu Qingzhen @.>; Mention @.> Subject: Re: [MIT-SPARK/TEASER-plusplus] Scaling seems doesn't work for python (#117)
- External Email -
Sure it returns an s value, however it is not correct. When the scaling ratio between the two pcds is quite big, it fails to address the registration and the scaling value doesn't really help much.
In any case, thank you for your time.
— Reply to this email directly, view it on GitHubhttps://github.com/MIT-SPARK/TEASER-plusplus/issues/117#issuecomment-1045245430, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AQXSV5LJGLD7CT7PELPPLK3U326QVANCNFSM5FB5LAGA. You are receiving this because you were mentioned.Message ID: @.***>
Notice: This email is generated from the account of an NUS alumnus. Contents, views, and opinions therein are solely those of the sender.
I see, did you test with an alternative feature descriptor, other than the fpfh, to see whether it performs better?
@ttsesm FPFH is not designed to be scale-invariant (see the original paper). Maybe try this: https://link.springer.com/article/10.1007/s00521-017-2964-1?