different result for diffierent order

[Wangweizhen-S]

I used my teeth photos (4 pictures), changed the input order, but why did the output result differ?

[yyfz]

Hi, Wangweizhen-S!

How do you change the input order? Did you input other modality (for example camera poses)?

[Wangweizhen-S]

Thank you for your response, I may wrote the wrong code before. However, when I use your officially example_mm.py just now with just two image folders where only the image order are not the same， the inconsistent result still happens.

Here are the paired result and input, and no other extra prior are given.

[yyfz]

Can you open the coordinate frame?

[Wangweizhen-S]

Can you open the coordinate frame?

please check

[yyfz]

Can you share me your input images? I can have a try.

[Wangweizhen-S]

Can you share me your input images? I can have a try.

images.zipSure. Thanks for your patience.

[yyfz]

Hi, in https://github.com/yyfz/Pi3/blob/main/pi3/utils/basic.py#L57 we resize all images based on the first image size. But the first image resolution is different in two folder. So the inference resolution is different. I try to fix the resolution to (686, 364) and the result is almost the same (the slight difference is because of machine calculation).

[Wangweizhen-S]

Hi, in https://github.com/yyfz/Pi3/blob/main/pi3/utils/basic.py#L57 we resize all images based on the first image size. But the first image resolution is different in two folder. So the inference resolution is different. I try to fix the resolution to (686, 364) and the result is almost the same (the slight difference is because of machine calculation).

Thank you for your response, I tried rezise the resolution and acquire the consistent result.
Regarding the fully permutation-equivariant architecture introduced in pi^3, I also have a question about its intrinsic benefits in highly controlled scenarios. Suppose we use a reference-anchored model (like VGGT) but guarantee that it is provided with a theoretically "optimal" reference frame. In such an ideal case, would pi^3 still yield better feature learning and reconstruction results? My intuition is that by eliminating the reference frame via permutation equivariance, pi^3 fundamentally changes the feature learning dynamics. Does this architecture encourage a more balanced, all-to-all feature interaction among all perspectives, rather than forcing each perspective to build a strong feature association solely with the reference frame?I am asking because I want to better understand exactly how the inductive bias of a fixed reference frame structurally limits feature learning, and whether the primary gain of pi^3—even when a perfect reference frame exists—comes from this enhanced global interaction. I would love to hear your insights on this.

[yyfz]

Yes, it would be better if we could manually compute a reference frame that minimizes the average error between each per‑frame point cloud and the ground truth (gt). Pi3 should have implicitly learned this reference frame.