Question about training-stage1

[LarkLeeOnePiece]

Based on my understanding, during the stage-1, you don't use the pose encoder to get the pose feature, but you randomly initialized geometry-feature and uv-coord-map using" uv_coord_map = getIdxMap_torch(torch.rand(3, posmap_size, posmap_size)).cuda()" and "geo_feature = torch.ones(1,self.net_parms.c_geom,self.model_parms.inp_posmap_size,self.model_parms.inp_posmap_size).normal_(mean=0., std=0.01).float().cuda()

self.geo_feature = nn.Parameter(geo_feature.requires_grad_(True))".

I am confused about the optimization, you said you just use these two tensor to capture the geometry feature and appearance feature. But how can you make sure they can capture the feature you want. Which operation and intuition that make it work as you want?
If we start from the posed SMPL and take the point from its surface to get the uv_coord_map, that's works. However, during the stage1, I can't see the relationship between the posed body points and uv_coord_map because you just generate the uv_map_coord randomly.

I would appreciiate it if you can give me some help.

[huliangxiao]

Hi, It's a good question.
The pose_featmap = None (

GaussianAvatar/model/avatar_model.py

Line 302 in d981c62

pred_res,pred_scales, pred_shs, = self.net.forward(pose_featmap=None,

) in Stage 1, the mean appearance features are learned by the geo_featrue.