Unable to reproduce the paper effect

[wangyuanbiubiubiu]

As mentioned in the paper, I input the 13-dimensional features (position, observation direction, variance scale, variance orientation), and encode the position and observation direction. Output features in 7 dimensions (delta_scale and delta_ orientation). But on my blur image, the effect is not as good as the original 3D GS, am I missing something?

[LZleejean]

Does your implementation work in Deblur-NeRF dataset?

[arvrschool]

I have the same problem. Have you solved the problem？I have seen the paper says 'The minima of these scaling factors (δr, δs) are clipped to 1 and element-wisely multiplied to r and s, ' I'm going to try.

[LiuYueXian123]

我也有同样的问题。你解决这个问题了吗？我看到论文说“这些缩放因子（δr，δs）的最小值被剪裁为1，并按元素乘以r和s，”我要尝试一下。

Please let me know if any progress has been made, I am also trying to reproduce the effect of the paper. Thank you so much！

[benhenryL]

Hi, thanks for your attention to our work and sorry for late reply.
The MLP takes 91-dimensional features (positionally encoded position (3 -> 21 dim), positionally encoded direction (3 -> 63 dim, we use 3-dim camera center as view direction to distinguish different views), scale (3 dim), and rotation (4 dim))
and output 7 dimensions, as you said.
Maybe positional encoding is not applied in your experiments and the outputs should be clipped to [1.0, 1.1].
New version of the paper with more details and codes will be uploaded soon.
Please let me know if you have other problems.
Thanks!