training process did not converge

[JunboLi-CN]

Dear Author(s),

Thank you for the great work! Could you please provide some example panorama testdata with which you have sucessfully done a reconstruction, because I have also tried your implementations but unfortunately the training process did not converge and I want to find out if my data have a problem.(If possible also with the output from opensfm) Thanks a lot!

Regards, Li

[inuex35]

Hello you can try the data from the URL. I use omnigs branch of diff gaussian rasterization. https://dtbn.jp/IeaSdxhe (URL will expire in a week)

These are training paramters. self.iterations = 30_000 self.position_lr_init = 0.000016 self.position_lr_final = 0.00000016 self.position_lr_delay_mult = 0.01 self.position_lr_max_steps = 30_000 self.feature_lr = 0.0025 self.opacity_lr = 0.01 self.scaling_lr = 0.001 self.rotation_lr = 0.001 self.percent_dense = 0.01 self.lambda_dssim = 0.2 self.densification_interval = 100 self.opacity_reset_interval = 3000 self.densify_from_iter = 500 self.densify_until_iter = 15_000 self.densify_grad_threshold = 0.0002 self.random_background = False

[JunboLi-CN]

Hello again, thank you for the reply. I have tried your data and it converged. But I got another question, are the outputs only the rendered panorama views from a fixed camera position or is like the original 3DGS which is a 3D scene in which I can move the camera position to see it from different perspective. I have used the SIBR_viewer to check the result but it looks reasonable only in the initial camera position, once I moved the camera position, it looks completely wrong. So I am wondering if the goal of your project is to first reconstruct a 3D scene and then render panorama images use the reconstructed scene or just render the panorama images directly as outputs?

[inuex35]

It does not happen from my environment. Could you tell me your environment? What branch do you use? Did you use my reconstruction.json? Reconstruction should be like this.

Sample image

reconstruction

https://github.com/user-attachments/assets/63211e79-51c7-4f22-904f-e869556377b8

[JunboLi-CN]

Hello, Thank you for the information. I have tried to reinstall the environment but the result looks still wrong.

I have used depth_normal_render and omnigs branch and yes I have used your reconstruction.json file and the parameters you provided. Do you have some idea about the possible reason for that result?

Screencast from 09-02-2024 10:50:31 AM.webm

[inuex35]

Thank you for the information. Could you try the main branch? Maybe depth normal render branch has some problems and I'm going to fix it.

[JunboLi-CN]

I have just tried the main branch and it works! Thank you for the reply and it helps a lot!