search

TencentARC / InstantMesh

InstantMesh: Efficient 3D Mesh Generation from a Single Image with Sparse-view Large Reconstruction Models
Apache License 2.0
3.41k stars 368 forks source link

Train on nerf of stage 1 with custom dataset #111

Open hopeliu20160622 opened 5 months ago

hopeliu20160622 commented 5 months ago

Hi, guys, i rendered my own rings dataset with 32 views , for every view, i have a RGBA image, where A means mask, normal image and depth image, and also saved the camera pos -- a 4*4 RT matrix. The size of the dataset is 8000. And i use the command to train: python train.py --base configs/instant-nerf-large-train.yaml --gpus 0,1,2,3,4,5,6,7 --num_nodes 1 After step = 7000, i checked the train images, as below, the third row, the two object has the same shape. Even i input a yellow duck multi-view images, the model output a ring like the third row. train_0001000 Can anyone help? Thanks. That's really weird.

bluestyle97 commented 5 months ago

I'm sorry I have no idea on this problem.

hopeliu20160622 commented 5 months ago

I'm sorry I have no idea on this problem.

Thanks ， can you share a training sample, I think there is something wrong with my render script, maybe i can compare the details to find the difference. Thanks. 504738295@qq.com

throb081 commented 5 months ago

@hopeliu20160622 hello,how did you render your datasets with 32 views?Use python code？i am interested in it .！！

ustbzgn commented 5 months ago

I'm sorry I have no idea on this problem.

Thanks ， can you share a training sample, I think there is something wrong with my render script, maybe i can compare the details to find the difference. Thanks. 504738295@qq.com

Hello，I am interested in render script for preparing training dateset, cloud you share me it? Thanks, fakezgn@126.com

guangxuwang commented 4 months ago

I'm sorry I have no idea on this problem.

Thanks ， can you share a training sample, I think there is something wrong with my render script, maybe i can compare the details to find the difference. Thanks. 504738295@qq.com

Hello，I am interested in render script for preparing training dateset, cloud you share me it? Thanks, gxwangmail@163.com

Mrguanglei commented 4 months ago

@hopeliu20160622 Hi, I already made the dataset, but I don't know how to add my own dataset in 'instant-nerf-large-train.yaml', can you share your config file? My dataset is of the following type for your convenience.Where in pose is the npz file of 32 viewpoints. In rgba there are 32 views (png type)，Looking forward to your reply ，mrguanglei@163.com 微信图片_20240714230249

zkaiWu commented 4 months ago

Hello, I'm also interested in the rendering scripts for rendering the training dataset. 869693441@qq.com, Looking forward to your reply.

denred0 commented 4 months ago

Can anybody post the rendering scripts here? It helps a lot of people. Thanks!

HaFred commented 3 months ago

Hi, guys, i rendered my own rings dataset with 32 views , for every view, i have a RGBA image, where A means mask, normal image and depth image, and also saved the camera pos -- a 4*4 RT matrix. The size of the dataset is 8000. And i use the command to train: python train.py --base configs/instant-nerf-large-train.yaml --gpus 0,1,2,3,4,5,6,7 --num_nodes 1 After step = 7000, i checked the train images, as below, the third row, the two object has the same shape. Even i input a yellow duck multi-view images, the model output a ring like the third row. train_0001000 Can anyone help? Thanks. That's really weird.

Hi @hopeliu20160622, may I know did you load in the openlrm checkpoint for training instantnerf? It seems like there are several tensor shape unmatching cases here.