search

SuLvXiangXin / zipnerf-pytorch

Unofficial implementation of ZipNeRF
Apache License 2.0
783 stars 85 forks source link

low quality #74

Open xiyufeng2 opened 11 months ago

xiyufeng2 commented 11 months ago

there is always too much floaters in rendered img. what should I do when I train my data? Thank you!

hermosaaurora commented 11 months ago

can you share the rendered img and the corresponding ground truth imgs?

cannedpeanutbutter commented 8 months ago

can you share the rendered img and the corresponding ground truth imgs?

i have the same problem! This is a frame from the video I rendered, render and the input image is here input The data is the video I took with my phone and extracted some of the frames from it.

I tried my own different data sets twice, training with batchsize=2048 due to OOM issues, but got similar results. Does zipnerf have any requirements for the input data set, or is it a problem with my training parameters? thank you！

xiyufeng2 commented 8 months ago

I find that the reconstruction quality is good in the outside environment whose view is very wide. But too many floaters in a room. Maybe the camera is too close to cause the floater?

cannedpeanutbutter commented 8 months ago

I find that the reconstruction quality is good in the outside environment whose view is very wide. But too many floaters in a room. Maybe the camera is too close to cause the floater?

Do you use colmap_and_resize.sh to generate poses_bounds.npy? I generate the bin file with colmap's simple radiation model rather than CV model and use imgs2poses in llff to generate the npy file. Is it the problem with the camera model in colmap? Please tell me all the specific configurations. Thank you!