Biased color at early stage of training / blurry foreground when fullly trained

[jingyibo123]

Kudos for the work!

When training the 360 dataset with the default 360_glo config using 4GPUs, I noticed two problems with the result:

1. During early stages of the training, the test image of foreground object shows weird color, For example the bicycle scene (I can't upload images...): The frame of the bicycle is went from red to gray to blue to white (still a bit blue)

2. When fully trained, the objects (chair and bicycle) are still blurry compared with the background, and both the chair and bicycle have a blue color shade..d

Could it be the view-dependent MLP? Thanks in advance!

[SuLvXiangXin]

@jingyibo123 I think this may results from the glo_vec proposed in paper which is used to scale and bias the bottleneck feature. If you get a weird result, you can try the default 360.gin instead.

[jingyibo123]

Indeed, thanks!