Anttwo / SuGaR

[CVPR 2024] Official PyTorch implementation of SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering

Home Page:https://anttwo.github.io/sugar/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Issues with single object (with background removed)

tanzim99901 opened this issue · comments

First of all, brilliant research work and a brilliantly documented codebase. I was able to get it up and running very easily.

I tested the code on various datasets and various scenes (even some of my own recorded videos) and it works perfectly. However, I was more interested in generating a mesh for single objects (with the background removed). I tried it on the "kitten" data (Link) from the Shelly dataset and it worked perfectly, as seen in the following figures:

kitten_01 kitten_02

However, when I tried it on one of my own objects with the background removed (a water bottle, Link, there are some noise (random black smokelike objects) around the bottle, as seen in the following figures:

bottle_01 bottle_02

Can you suggest how I can try solving this issue? Like what parameters can I try changing to remove the background noise? I am assuming since the code worked for the "kitten" object, it should work for other objects as well.

I look forward to hearing back from you. And again, thank you for the great work!

P.S.:
I ran the following commands:

For "kitten"

  1. Kept all images: SuGaR -> kitten -> input
  2. COLMAP estimation: python gaussian_splatting/convert.py -s kitten
  3. Vanilla Gaussian Splatting: python gaussian_splatting/train.py -s kitten --iterations 7000 -m kitten_GS_out -r 1
  4. SUGAR refinement: python train.py -s kitten -c kitten_GS_out/ -r "density" --refinement_time "long" --high_poly True

For "bottle"

  1. Kept all images: SuGaR -> bottle -> input
  2. COLMAP estimation: python gaussian_splatting/convert.py -s bottle
  3. Vanilla Gaussian Splatting: python gaussian_splatting/train.py -s bottle --iterations 7000 -m bottle_GS_out -r 1
  4. SUGAR refinement: python train.py -s bottle -c bottle_GS_out/ -r "density" --refinement_time "long" --high_poly True

For both cases, I tried with both vertices_density_quantile = 0.1 and vertices_density_quantile = 0. and got the same results.