not clear img results
Calsia opened this issue · comments
hi, thks for this great work for large scene.
i uesd it for my own dataset (500 pics), after training 500k, i get the img from your 'eval_result_rgbs' from the scritpt 'eval.py'.
but the result rgb img is not clear enough, do you have any suggestion?
any reponse is helpful. thanks!
by the way, there is 2 param that i am not very clear.
one is ray_altitude_range, you explained it in other issue "The ray_altitude_range parameter is used to truncate ray sampling near ground level. You generally want to set the first value to be the height of the largest object in the foreground that you care about and the second to be close to ground level. " but i don't understand why the hight can be negative in some your config，and the second value is much bigger than the first ，does that mean the ground is above the object？
the other one is pose_scale_factor， i set it be 1000 to make my camera pose in unit （the center of all my camera is [ -91.4179159 -121.95918623 952.39327828]）， was it right？
the world coordinate system used in mega-nerf is (down, right, backwards), so negative values are higher altitude-wise than positive values. In retrospect I would have used a more intuitive world coordinate system, but hindsight is 20/20 :)
I'd double-check that all of your camera poses are within the [-1, 1] range, so if your origin is centered that looks reasonable.
In general regarding bluriness I'd make sure that your camera poses are accurate, and use something like PixSFM if you can (we personally saw a huge difference between PixSFM and commercial drone mapping software when running experiments for the paper)
