SegAnyGAussians icon indicating copy to clipboard operation
SegAnyGAussians copied to clipboard

About segmentation results

Open songyuri opened this issue 1 year ago • 10 comments

Hello, thanks to your help, I was able to successfully use the GUI to check the segmentation. However, I'm not sure how to review the segmentation results, so I'm posting another issue.

Currently, in the GUI, I've selected objects using the multiclick mode, and they are highlighted in red. I wanted to segment these objects, so I clicked segment3D, but nothing changed. Is this normal? Then, I clicked save as to save the results in segmentation_res as a .pt file. image image

Additionally, I ran the command: python render.py -m output/squirrel/ --precomputed_mask segmentation_res/precomputed_mask.pt --target scene --segment

and it generated images in the train/ours_30000/ directory, but these images are of the entire scene, not the segmented results. There are no results in the mask. A new folder named test was also created, but it doesn't contain any results.

image this is train folder

image And this is test/gt/, but there is nothing in. mask and renders are same state.

I want to view the segmented results in 3D, but I'm not sure how to proceed. I thought I might need to use Jupyter, but when I tried, I encountered the following error in the first cell: ModuleNotFoundError: No module named 'utils.contrastive_decoder_utils'. image

Could you tell me what I have to do? Thank you so much!

songyuri avatar Jun 21 '24 15:06 songyuri

Hi, if you want to conduct 3D segmentation, you need to adjust the hyper parameters first, i.e., the scale and score thresh. In default, the score thresh is 0.0, where all 3D Gaussians will be reserved, that's why it looks like nothing happened.

For the utils.contrastive_decoder_utils, you can just remove it.

Jumpat avatar Jun 21 '24 15:06 Jumpat

Hi I removed that code but another error was occur image Could you explain about this?

songyuri avatar Jun 26 '24 08:06 songyuri

The clip_features is used for language-driven segmentation. Check the section of Data and Model Preparation and make dataset.need_features = True.

Jumpat avatar Jun 27 '24 06:06 Jumpat

Is the GUI only nsed on jupyter-notebook? I run the saga_gui.py on VSCode using server, and seems it can not work

Zqzne avatar Dec 02 '24 09:12 Zqzne

I run the saga_gui.py 1

RuntimeError: expected mat1 and mat2 to have the same dtype, but got: float != c10::complex

Could you tell me what I have to do? Thank you so much!

feng-xiyun avatar Dec 04 '24 01:12 feng-xiyun

Is the GUI only nsed on jupyter-notebook? I run the saga_gui.py on VSCode using server, and seems it can not work

The GUI cannot run on a remote server. You need to run it locally.

Jumpat avatar Dec 04 '24 02:12 Jumpat

I run the saga_gui.py 1

RuntimeError: expected mat1 and mat2 to have the same dtype, but got: float != c10::complex

Could you tell me what I have to do? Thank you so much!

This seems strange. Looks like CUDA version error, or torch version error, or something else. We haven't met such error yet.

Jumpat avatar Dec 04 '24 02:12 Jumpat

 找到方法了, 将saga.gui的594行修改为 sem_transed = sems.float() @ self.proj_mat.float()即可解决,谢谢 @Jumpat

feng-xiyun avatar Dec 04 '24 02:12 feng-xiyun

Is the GUI only nsed on jupyter-notebook? I run the saga_gui.py on VSCode using server, and seems it can not work

The GUI cannot run on a remote server. You need to run it locally.

All right, thanks!!!

Zqzne avatar Dec 04 '24 07:12 Zqzne