on my own data

[Hxinyue]

Thanks for your work! I only have points in .ply format, how can I retrain or generate them directly?

[tangjiapeng]

Hi, you can put your .ply files under data/demo/synthetic_room_dataset, and then run

python generate_optim_scene.py configs/pointcloud/demo_syn_room.yaml

[Hxinyue]

Thank you very much for your reply! Do I need to change my ply format data to the same style as "syntactic_room_dataset"?

[tangjiapeng]

You need to normalize the point clouds to [-1, 1].

[zhao-you-fei]

@tangjiapeng 您好，请问我在自己的数据集（室内场景）上没有取得很好地效果，可以修改那些参数来增强效果。

[Hxinyue]

You need to normalize the point clouds to [-1, 1]. Hello, I am trying to run "python generate_optim_scene.py configs/pointcloud/demo_syn_room.yaml" and an error message is reported as follows. May I know where the problem is with me?

[zhao-you-fei]

非常希望和您讨论学习，我的问题请详见附件

------------------ 原始邮件 ------------------ 发件人: "tangjiapeng/SA-ConvONet" @.>; 发送时间: 2023年4月13日(星期四) 中午1:42 @.>; @.**@.>; 主题: Re: [tangjiapeng/SA-ConvONet] on my own data (Issue #11)

You need to normalize the point clouds to [-1, 1]. Hello, I am trying to run "python generate_optim_scene.py configs/pointcloud/demo_syn_room.yaml" and an error message is reported as follows. May I know where the problem is with me

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

[Hxinyue]

@tangjiapeng Hello, I understand that the loss nature of UCE adds a layer of RulU to the output. What is the relationship between this design and normal line? And whether the "we force the observed surface P to align with the 0.5 level set of occupancy field, and the signed occupancy values of non-surface points to be either 0 or 1"in the article will lose information?

[tangjiapeng]

The reason why we don't use common BCE is that we don't know the inside/outside information due to the lack of normals. Based on the pre-trained geometric priors, we can still optimize the occupancy fields, by enforcing the occupancy values of non-surface points as either 0 or 1. The prior initialization can provide some constraints that the points outside the surface would not be pulled to the inside during the optimization of UCE.

[Hxinyue]

The reason why we don't use common BCE is that we don't know the inside/outside information due to the lack of normals. Based on the pre-trained geometric priors, we can still optimize the occupancy fields, by enforcing the occupancy values of non-surface points as either 0 or 1. The prior initialization can provide some constraints that the points outside the surface would not be pulled to the inside during the optimization of UCE.

Hello, may I ask if this zero setting operation will cause loss of output information?

[zhang-zhe3203]

您好，非常感谢您的工作，请问我只有一个.ply文件，我想用模型将.ply直接生成mesh并输出，应该怎么做

[tangjiapeng]

you can generate the npz file from ply file by "python scripts/dataset_matterport/make_cropscene_dataset2.py --in_folder data/demo/Matterport3D_processed --out_folder data/demo/Matterport3D_processed_normalize --do_norm".

[zhang-zhe3203]

您可以通过“python scripts/dataset_matterport/make_cropscene_dataset2.py--in_folder data/demo/Materport3D_processed--out_folder-data/demo/Materport 3D_pprocessed_normalize--do_norm”从ply文件生成npz文件。

Thank you very much for your reply. I have successfully run it according to your guidance！

[Hxinyue]

Thank you very much for your work, there is no problem at present, it will be closed