Optimizing the texture coordinates of a mesh using meshlab.

[jmespadero]

This script was created as side result of this question in stackOverflow about how to edit and save texture images in Meshlab, which can be summarized as:

Q: "Can I export an updated and much smaller texture image file from within MeshLab?"
A: "No, you can't. But..."

I have seen many projects addressing this problem, like this nice Pythagoras Bust model originally published by Geoffrey Marchal.

which uses this huge 8192x8192 texture:

where clearly only the bottom third of the image pixels are being used as texture, so it could be stored in just a 8192x2882 image without any loss.

This is not a specific problem of this model, but a quite common problem generated by some photogrammetry software, so there is lots of models that has texture images like that.

I have written some small script with pymeshlab that can do some work done (basically, compute the bounding box of the texture coordinates, crop the original texture image to contain just the referenced bounding box, and then remap the old texture coordinates into the new texture).

The good news are: Here is the code...

#!/usr/bin/env python3
import pymeshlab as ml
import numpy as np
import sys

try:
    # Get input filename from argument 1
    fn=sys.argv[1]
except:
    print(f'Usage: {sys.argv[0]} input_filename')
    sys.exit(-1)

# Load input mesh and dump some info
ms = ml.MeshSet()
ms.load_new_mesh(fn)
m = ms.current_mesh()
print('Input mesh:', fn, 'vertex_number=',m.vertex_number(), 'face_number=', m.face_number())
for t,img in m.textures().items():
        print ('- Texture:',t, img.width(),'x',img.height() )

# Get filename of first texture
texfn=next(iter(m.textures()))

# Ensure we have wedge texture coordinates
try:
    ms.apply_filter('compute_texcoord_transfer_vertex_to_wedge')
except:
    pass

# Analyse texture coordinates to get the referenced area of the image
wt = m.wedge_tex_coord_matrix()
bbmin = np.amin(wt, axis=0)
bbmax = np.amax(wt, axis=0)
print('Input texture real coordinate range: ', bbmin, "to", bbmax)

# Clamp values to avoid too small changes that introduce changes with no visible gain
if bbmin[0] < 0.05 : bbmin[0] = 0.0
if bbmin[1] < 0.05 : bbmin[1] = 0.0
if bbmax[0] > 0.95 : bbmax[0] = 1.0
if bbmax[1] > 0.95 : bbmax[1] = 1.0
print('Input texture clamped coordinate range: ', bbmin, "to", bbmax)

# Compute cropped texture size
w=bbmax-bbmin
newOrigin=[int(bbmin[0]*img.width()), int(bbmin[1]*img.height())]
newSize=[int(w[0]*img.width()), int(w[1]*img.height())]
print('New texture image size:', newSize)

# Automatically create cropped and resized new texture image, using PIL/Pillow
from PIL import Image
tex = Image.open(texfn)
crop=tex.crop((newOrigin[0], img.height()-newOrigin[1]-newSize[1], newOrigin[0]+newSize[0], img.height()-newOrigin[1]))
crop.thumbnail((2048, 2048)) #This is lossy resize. Remove this line if you want lossless
crop.save(f'new_{texfn}')
print(f'new texture image file saved as new_{texfn}')

# I prefer not to automatically resize image with PIL, because sometimes I do some manual edition,
# and sometime textures are not in the same folder, etc...
print('You can also use an external program to crop and resize input image file. For example:')
print(f'convert {texfn} -gravity South -crop {newSize[0]}x{newSize[1]}+{newOrigin[0]}+{newOrigin[1]} -resize "2048x2048>" new_{texfn}')

# Rescale texture coordinates from input domain to [0..1]
wt2 = (wt-bbmin) / w

# Create new mesh with new texture coordinates and add it to MeshSet
#m2=ml.Mesh(m.vertex_matrix(), m.face_matrix(), m.edge_matrix(), m.vertex_normal_matrix() )
m2=ml.Mesh(m.vertex_matrix(), m.face_matrix(), m.edge_matrix(), m.vertex_normal_matrix(), w_tex_coords_matrix=wt2 )
ms.add_mesh(m2, "reTextured")
ms.set_texture_per_mesh(textname = f'{texfn}')

# Export as output.ply
print('Export new mesh as output.ply')
ms.save_current_mesh("output.ply")

The bad news are: This script depends of a feature introducted in a recent comint f7c346c to fix issue #297 , so if you are reading this in june of 2023 you need to wait until next release of PyMeshlab (or download the current source code of pymeshlab from github).