0
$\begingroup$

I have a huge dataset of 3D point clouds (each point consists of X,Y,Z coordinates) and another dataset with keypoints (also X,Y,Z) which lie on quite recognizable structures in the point cloud. As a human it is pretty easy to find these keypoint given the according point cloud.

There are 6-24 keypoints for each point cloud which may have up to 100.000 points but can be downsampled to fit into memory while processing.

My goal is to train a supervised deep learning model to detect/generate these keypoints in new point clouds. These have to be new synthetic points since the keypoints are most likely not in my point cloud.

Previously i tried to train a regression model using PointNet++ semantic segmentation. I generated weights for each point in the input point cloud according to its proximity to the next keypoint. For new point clouds the weights were predicted pretty well but it is very difficult to regenerate the synthetic keypoints from these weights since the area around the keypoint can be very sparse.

Essentially i am trying to use a model like this RSN on 3D point clouds to detect synthetic keypoints in the data. I thought about using object detection models like VoxelNet but these detect bounding boxes and i want to get a single predicted point as a result.

I am also wondering if its worth looking into GANs but i dont have experience with these models and until now i only found models who generate a whole new point cloud instead of a single points inside a given point cloud.

I would love to hear some ideas from you to point me in the right direction.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

The previous milestone of your question is this*

You can use data stack that it has low resolution as S(Yn) of stream from supervised network for verifying. No data transfer loss with this rule:

S(Xn) ≥ S(Yn) > S(Xn)/2

Each time you change tether provider that network model before transform data stacks for the default stream or more data stack may be required because less finite sub-states in the supervised network will not be a factor in diagnosis.

There is an invert ratio between signal complexivity of X and process accuracy in S(Yn) ≤ S(Xn)/2 rule.

Improve this answer
$\endgroup$
5
  • $\begingroup$ As it’s currently written, your answer is unclear. Please edit to add additional details that will help others understand how this addresses the question asked. You can find more information on how to write good answers in the help center. $\endgroup$
    – Community Bot
    Commented Mar 9, 2023 at 16:41
  • $\begingroup$ Unfortunately I do not understand your answer either. Your linked post is also inaccessible.. $\endgroup$
    – nmb
    Commented Mar 16, 2023 at 11:57
  • $\begingroup$ Hmm... Link is broken. It fixed. If can help you this, <i>Y<sub>n</sub></i> is your check data in one process time. Obtain check data by new tether before decode this by referencing first tether in every process time. <i>Y</i> is your check data stack and syntesis doors by this. This dynamic to be more successful than using static tethering methods because doors have exceptions of finite sub processes that they can not detected approximately. $\endgroup$
    – fkybrd
    Commented Mar 16, 2023 at 16:14
  • $\begingroup$ You can use mathjax on this site! $\endgroup$
    – nbro
    Commented Apr 17, 2023 at 12:31
  • $\begingroup$ Rules are not providing my requirements on this site. You can delete those $\endgroup$
    – fkybrd
    Commented Apr 19, 2023 at 8:59

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .