Does the metric support input with mesh and texture?

[jiangzhongshi]

Hi,

Thanks for the great work and code.
I want to estimate the quality of a new mesh, with obj and texture file, is there an instruction to input then and go through the renderings to get the final score?

Thanks,
Zhongshi

[glavoue]

Dear Zhongshi,
The metric takes as input local patches. So you have to make your renderings on your own and create your patches (64*64) before calling the metric.
The global quality score of the model is computed as the average of the local patch qualities.
I hope it's clear!

Guillaume

[jiangzhongshi]

Hi Guillaume,

Thank you so much for the prompt response!

Do you have code pointers to make the renderings and/or any specification and recommendations for material/camera parameter, image count and camera positioning?

Thanks,
Zhongshi

[glavoue]

Here is what we did in our experiments, and that you can follow:

The rendering was done using Unity with a Lambertian material and mipmapping activated (650 × 550 image size). We used a directional light coming from the top right.

For computing the metric, we patchified the images into small overlapping patches of size 64 × 64 sampled every 32 pixels., and we removed patches containing less than 65% stimulus information.

For the number of views, we selected either the most relevant view or computed 4 views around the object (see sec. 5.5 of the paper)

Below the parts of the paper that describe that information in details:

“We rendered the dynamic test stimuli in a neutral room (light gray walls), without shadows and under a directional light coming from the top right of the room. All models were approximately the same size and rendered with a lambertian material; mipmapping was activated”

“The image size, 650 × 550, is the video resolution of the stimuli seen by the participants in the subjective experiment. We divided (patchified) the images into small overlapping patches of size 64 × 64 sampled every 32 pixels. We removed patches containing less than 65% stimulus information (i.e., the percentage of background pixels in the patch is greater than 35%). We got an average of 60 patches per stimulus."

Sec 5.5. : "For each stimuli, we thus generated 4 snapshots taken from 4 camera positions regularly sampled on its bounding box and prepared the data as for the above network. We used the same training and testing set as our representative fold and used the same training parameters, randomly sampling 𝑁𝑝 = 300 patches from all possible viewpoints"