Although pure neural network (SIREN) demonstrated effective regularization, it tended to oversmooth the reconstruction results.

[+] Fine regularization

[-] Oversmoothing


While Poisson reconstruction exhibited a generally satisfactory level of reconstruction quality, it was prone to floating artifacts and exhibited limited adaptability to sparse regions.

[+] Good global quality

[+] Adaptive to Input

[-] Strong local artifacts in certain scenes


The combination of neural network and octree (SHINE) yielded impressive overall quality that adapted well to the input. However, the resulting reconstruction still exhibited notable strong local artifacts.

[+] Overall good reconstruction result

[-] Floating artifacts

Future Works

  • Exploit sequential information for better empty space
    • Trim away the vertices in the reconstructed global mesh where the density is below a certain threshold
  • Extend our method and test them on outdoor scenes