Inaccurate prediction of normal map

[Moreor1ess]

I found that when predicting normal maps, especially on the back, it is easy to predict leg bending and tiptoeing, which leads to the generated model appearing suspended or unable to stand properly. But if the camera position can be lowered to avoid this situation, I would like to know how to correct this or how to train my own normal prediction model

[YuliangXiu]

This comes from the inaccurate HPS (human pose estimator) instead of the normal estimator. And the normal estimator is conditioned on HPS results as well.

• Something strange in the expoting obj file #5
• Why bent legs appears given the normal standing human photo? ICON#68

[Moreor1ess]

Thank you for your reply！

[YuliangXiu]

@Moreor1ess

The issue of bending legs get resolved finally, please check out tricks.md/bending-legs.

[Moreor1ess]

Thank you for your reply, it will help me a lot

[Moreor1ess]

After I applied your modified code, I had foot reconstruction errors on various pictures, but the reconstruction results of other parts were very similar. I would like to know if you have any ideas.

[YuliangXiu]

I guess this is related to the disparity in cameras between synthetic data (rendering with orthogonal camera) and real data (which generally uses a perspective camera).