423A35C7/pytorch3d
1
0
Fork 0
mirror of https://github.com/facebookresearch/pytorch3d.git synced 2025-08-03 04:12:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

(bug) Fix exception when creating a TextureAtlas

Browse Source 
Summary: Fixes GitHub issue #751. The vectorized implementation of bilinear interpolation didn't properly handle the edge cases in the same way as the `grid_sample` method in PyTorch.

Reviewed By: bottler

Differential Revision: D30684208

fbshipit-source-id: edf241ecbd72d46b94ad340a4e601e26c83db88e
This commit is contained in:
Nikhila Ravi 2021-09-01 09:08:12 -07:00 committed by Facebook GitHub Bot
parent 835e662fb5
commit fc156b50c0
2 changed files with 8 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
pytorch3d/io/mtl_io.py
View File

@ -299,7 +299,7 @@ def make_material_atlas(
# bi-linearly interpolate the textures from the images # bi-linearly interpolate the textures from the images
# using the uv coordinates given by uv_pos. # using the uv coordinates given by uv_pos.
textures = _bilinear_interpolation_vectorized(image, uv_pos) textures = _bilinear_interpolation_grid_sample(image, uv_pos)
return textures return textures
@ -311,11 +311,14 @@ def _bilinear_interpolation_vectorized(
Bi linearly interpolate the image using the uv positions in the flow-field Bi linearly interpolate the image using the uv positions in the flow-field
grid (following the naming conventions for torch.nn.functional.grid_sample). grid (following the naming conventions for torch.nn.functional.grid_sample).
This implementation uses the same steps as in the SoftRas cuda kernel This implementation uses the same steps as in the SoftRasterizer CUDA kernel
to make it easy to compare. This vectorized version requires less memory than for loading textures. We are keeping it for reference to make it easy to
compare if required.
However it doesn't properly handle the out of bound values in the same way as
the grid_sample function does with the padding_mode argument.
This vectorized version requires less memory than
_bilinear_interpolation_grid_sample but is slightly slower. _bilinear_interpolation_grid_sample but is slightly slower.
If speed is an issue and the number of faces in the mesh and texture image sizes
are small, consider using _bilinear_interpolation_grid_sample instead.
Args: Args:
image: FloatTensor of shape (H, W, D) a single image/input tensor with D image: FloatTensor of shape (H, W, D) a single image/input tensor with D

BIN
tests/data/test_texture_atlas_8x8_back.png
View File

Binary file not shown.
Side by Side Swipe Overlay

Before

Width:  |  Height:  |  Size: 31 KiB

After

Width:  |  Height:  |  Size: 31 KiB