|
--- |
|
license: mit |
|
datasets: |
|
- ShapeNet/ShapeNetCore |
|
title: Transformer-Based Point Cloud Denoising/Completion Models |
|
--- |
|
|
|
Four variants of the model architecture were trained on a random subset (50%) of ShapeNetCore. |
|
The models are based on two types of data augmentations: random noise perturbations (Gaussian normal) and point deletion (using uniform random sampling). |
|
The models are as follows: |
|
|
|
|
|
| Model | noise strength (%) | points removed (%) | |
|
| :---------------- | :------: | :----: | |
|
| pc_denoiser_1 | 1 | 0 | |
|
| pc_denoiser_2 | 1 | 50 | |
|
| pc_denoiser_3 | 2 | 0 | |
|
| pc_denoiser_4 | 2 | 50 | |
|
|
|
|
