search

mlvlab / UP-NeRF

Official Implementation (PyTorch) of "UP-NeRF: Unconstrained Pose-Prior-Free Neural Radiance Fields", NeurIPS 2023
MIT License
22 stars 3 forks source link

Good NVS Quality #5

Closed TwiceMao closed 4 months ago

TwiceMao commented 4 months ago

@cmh1027 I found that your method is not very effective in camera pose optimization when there are not many angles of view, but the synthesis effect of new angles is also good. Why is this? Is it related to Appearance Embedding ?

cmh1027 commented 4 months ago

appearance embedding is for color-inconsistency in wild images so it can not recover poor NVS quality from poorly optimized pose. There can be two candidates for the situation. First, estimation of poses is not correct due to wrong implemetation somewhere in our code (I'm not sure about where it would be). Second, near and far parameters are set to 0.1, 5.0 each. Maybe far parameter is too high for your dataset (I experiened similar situation in the past for Lincoln scene in Phototourism) You can scale it down to like 2.0.

TwiceMao commented 4 months ago

Thx for your kind reply~