File size: 2,414 Bytes
4287e47 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 |
ModelNet_Splats release (Fall 2024)
Each zip file contains ply files where each Gaussian is encoded as a vertex with custom vertex attributes. This ply format is commonly used for Gaussian splats and can be viewed using [online viewer](https://playcanvas.com/supersplat/editor/).
To open the .ply files, you can use the following python code:
```python
from plyfile import PlyData
import numpy as np
gs_vertex = PlyData.read('ply_path')['vertex']
### load centroids[x,y,z] - Gaussian centroid
x = gs_vertex['x'].astype(np.float32)
y = gs_vertex['y'].astype(np.float32)
z = gs_vertex['z'].astype(np.float32)
centroids = np.stack((x, y, z), axis=-1) # [n, 3]
### load o - opacity
opacity = gs_vertex['opacity'].astype(np.float32).reshape(-1, 1)
### load scales[sx, sy, sz] - Scale
scale_names = [
p.name
for p in gs_vertex.properties
if p.name.startswith("scale_")
]
scale_names = sorted(scale_names, key=lambda x: int(x.split("_")[-1]))
scales = np.zeros((centroids.shape[0], len(scale_names)))
for idx, attr_name in enumerate(scale_names):
scales[:, idx] = gs_vertex[attr_name].astype(np.float32)
### load rotation rots[q_0, q_1, q_2, q_3] - Rotation
rot_names = [
p.name for p in gs_vertex.properties if p.name.startswith("rot")
]
rot_names = sorted(rot_names, key=lambda x: int(x.split("_")[-1]))
rots = np.zeros((centroids.shape[0], len(rot_names)))
for idx, attr_name in enumerate(rot_names):
rots[:, idx] = gs_vertex[attr_name].astype(np.float32)
rots = rots / (np.linalg.norm(rots, axis=1, keepdims=True) + 1e-9)
### load base sh_base[dc_0, dc_1, dc_2] - Spherical harmonic
sh_base = np.zeros((centroids.shape[0], 3, 1))
sh_base[:, 0, 0] = gs_vertex['f_dc_0'].astype(np.float32)
sh_base[:, 1, 0] = gs_vertex['f_dc_1'].astype(np.float32)
sh_base[:, 2, 0] = gs_vertex['f_dc_2'].astype(np.float32)
sh_base = sh_base.reshape(-1, 3)
```
Note that more details regarding <u>PSNR</u>, <u>SSIM</u>, <u>LPIPS</u>, <u>FILE_SIZE</u> and <u>Number of Gaussians</u> can be find in [summary](ShapeSplatsV1_Qualitative_Results.json). To showcase the difference of point cloud in spatial distribution, we report Jensen–Shannon Divergence (JSD) and Maximum Mean Discrepancy (MMD) between gaussian splats and ShapeNetCoreV1 point clouds.
More details for the pretraining method of ShapeSplats can be find in [Gaussian-MAE](https://unique1i.github.io/ShapeSplat/)
Last updated: 2024-09-05
|