How to improve result?

[hardikdava]

Hello @wanmeihuali , I have started doing some experiments with the code. The implementation is really good and clean, I like it. I have couple of questions regarding result quality.

• Result is a bit blur and missing some fine details.
• Because of black ground PSNR and SSIM score look good.

• Left = Rendered View, Right = Groundtruth View

  

Questions:

• Which parameter affeacts for image sharpness?
• Is it possible to know a little description of parameters?
• Can we use RayTune or Optuna tool to get tunned parameters?

P.S. I am ready to provide a PR if something is improving the result.

[wanmeihuali]

Hi @hardikdava , there could be a lot of parameters that affect image sharpness, Given this is your private dataset, I can only provide some ideas:

• input camera pose; inaccurate camera pose reduces the quality of training.
• input images; make sure your image is undistorted and the resolution isn't too low.
• input pointcloud; try to provide a more dense input point cloud, or choose a more radical densify policy.
• densify policy:

  adaptive-controller-config:
  # The threshold on the gradient of point position in the camera plane for all affected pixels
  # decrease it->more points, slower train/render speed, better quality
  densification-view-space-position-gradients-threshold: 6e-6 
  # The frequency to split/clone points, shall adjust base on dataset size
  num-iterations-densify: 100
  # The frequency to reset all point opacity, if you have more images, maybe you shall increase this value, otherwise points may still be transparent. Also, please note that rendering quality will drop dramatically in the few iterations after reset.
  num-iterations-reset-alpha: 4000

  About autotuning tools: sure, you can try to find the best parameters in the YAML config.

[hardikdava]

@wanmeihuali Thanks for your reply. I will get back to you on this with the result.

[hardikdava]

@wanmeihuali I have better results after tunning the parameters like you suggested.

[jb-ye]

@hardikdava could you try using white background for the segmented image? Note sure if those wash-out effects still exist with this.