Interpretations of One Ablation

[GloryyrolG]

Hi @tiangexiang et al.,

Thx for ur great constant contribution in this field. May I ask in Sec. 4.6 & Fig. 8 1st row, it states:

... by inheriting the two-stage parameterization but optimizing the rendering of the foreground human on the visible parts only [46] while maintaining the same optimization objectives.

if this is the case, then why it will lead to bad rendering results in Fig. 8? I am a bit confused and may miss sth. Is that becuz there are no sufficient frames where the problematic part is visible? Many thx in advance for any help. Rgds,

[tiangexiang]

Good question! My interpretation is that:

1. Human's appearance is conditioned on SMPL poses [vid2avatar], such that for different poses, the rendering can be different, even though for ray samples that are sampled at the same position in the canonical space.
2. For this particular pose, only a part of human is visible, therefore only the rays at the corresponding visible area are casted and sampled during training.
3. Based on the above two points, we can see that the rendering of human appearance depends on two factors: both pose and ray sample positions. However, such combinations at he invisible area are never seen during training, making the rendering of invisible part funny.

[GloryyrolG]

Hi @tiangexiang ,

Thx for ur instant reply first. I still do not fully understand.

1. For 1st point, do u mean by the view-dependent effect which affects the alpha composition/integral along different rays?

2. I understood 2nd point:)

3. First ignoring the logic I may not understand correctly, do u imply those invisible areas are not observed in any training frames? E.g., this part is always occluded for people always hold the object. Though I do not think that is the case. Cuz if so, Wild2Avatar is not able to trace back part appearance either.

Thx for help.