Fangneng Zhan

> There are two problems, 1. The 3D designed by the Alibaba designer should be rotated by 90 (or -90) in Z axis, I manually adjust it for my experiments....

Hi, I just find including sigmoid will make the network more difficult to converge during training.

It seems to be caused by the image size, you may check it. The default size of panorama in my work is 256 * 128.

> Yes, it requires some preprocessing as described in "deep parametric indoor lighting estimation". I directly got the preprocessed data by contacting the author.

The warping operation is quite trivial, we try to implement it but give up soon. Later, we directly obtained the warped image from author of [1].

You can divide a validation set, and compute the L1 distance between the gaussian maps reconstructed by predicted parameters and gt parameters as the validation accuracy. Then the best regression...

> TKS very much for your helpNow,I have some problems?Please guide me! > 1、train.py script produce the summary.jpg，the left is original image,Is the right predicted map? the right image is...

> help me quantitate the result?  Here are some expected training results:   

> Hi, fnzhan: the trained datas of the geprojector also divided as the same of regression process? Sorry for the late reply, there is too many things and deadlines. Definitely,...

The procedure: replacing the illumination map, rendering the inserted object, saving the rendered object image as .png (rgba), putting the rendered rgba image on the background, i.e., final insertion result....