search

nywang16 / Pixel2Mesh

Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images. In ECCV2018.
http://openaccess.thecvf.com/content_ECCV_2018/papers/Nanyang_Wang_Pixel2Mesh_Generating_3D_ECCV_2018_paper.pdf
Apache License 2.0
1.64k stars 293 forks source link

Could you tell your strategy of sampling point clouds? #38

Open xiaomingjie opened 5 years ago

xiaomingjie commented 5 years ago

HI @nywang16 I looked into your ShapeNetTrain dataset and found that the point numbers of each model are not the same. It seems that you didn't mention your strategy of sampling point clouds in your paper. Could you tell me about that?

xiaomingjie commented 5 years ago

I think I've found your strategy in generate_data.py. But I still wonder that why you use different point numbers to sample a mesh instead of a fixed number like 16384 in PSGN. Especially, I noticed that some GTs even got less than 2466 points which are the numbers of your predicted mesh. Would that be a problem?

walsvid commented 4 years ago

Our sampling strategy is to use Poisson-disk sampling. Since the fast version code of the Poisson-disk sampling cannot control how many points will eventually be sampled, the GT points of different data are not the same. This is indeed a potential problem in data pre-processing. We have a solution to this problem and we will update the solution in the future.

zhiwenfan commented 4 years ago

Our sampling strategy is to use Poisson-disk sampling. Since the fast version code of the Poisson-disk sampling cannot control how many points will eventually be sampled, the GT points of different data are not the same. This is indeed a potential problem in data pre-processing. We have a solution to this problem and we will update the solution in the future.

Hi, are you willing to release the new sampling strategy recently?

Libuyan00 commented 1 year ago

Our sampling strategy is to use Poisson-disk sampling. Since the fast version code of the Poisson-disk sampling cannot control how many points will eventually be sampled, the GT points of different data are not the same. This is indeed a potential problem in data pre-processing. We have a solution to this problem and we will update the solution in the future.

Hello, I am making my own data set now. When sampling the model points, I found that the sampling points are rarely less than 7000 points. Is there any new sampling strategy or solution? thank you