[ ICCV 2021 Oral ] Our method can estimate camera poses and neural radiance fields jointly when the cameras are initialized at random poses in complex scenarios (outside-in scenes, even with less texture or intense noise )
), the size is enforced to be proportional to the original size which is 1600x1200 for DTU, thus [500, 400] does not work. Should the size be [400, 300] instead? I set the image size to be [400, 300] in my following experiments.
I got pretty blurry synthesis results on DTU, for example, on scan63, after 30K iters, I got the following results,
on scan4, after 30K iters,
What are the pose estimation scores (rotation and translation errors) on DTU dataset?
There was an error while trying to return the code to my own blender data similar to nerf_synthetic data. It was confirmed that all the results derived from phase ABAB (from 12k) did not converge and splashed. (continued until phase b) Do you know why this is happening? I trained by python train.py ./config/blender.yaml --data_dir PATH/TO/DATASET and the data has same forms just like the nerf_synthetic data.
Thank you for the code.
Hello authors,
I have read the GNeRF paper recently and try to re-product the results on blender hotdog and lego dataset, with released code in this repository and default settings. But I find that the output gif result is all white after at least 30k training iterations.
I have also read the previous issues proposed previously here. The author say that it happens that the GAN part training fails and leads to all-white results. I wonder whether it is normal for the GNeRF to fail on GAN training?
I have some problem for the network result.
In the end, the generator and discriminator look like convergence. But the result gif is all NONE. (rgb gif is all white, depth gif is all Black)
I am so confuse about that.
What I have changed before training this network is only changed the batch size from 12 to 6. Does this change make this err? :)
ps: I use the blender drums dataset for training.
one GPU 2080Ti
Training for almost 4 days
Do you have any advice on the loss function, which do you prefer, wgan loss or standard loss?
I find in released code that the standard loss is the default option for you, have you test wgan loss ?
Thanks for the interesting work and releasing the code.
I was wondering about the advice that you've put in the readme on training with our own data.
Assuming that I only have an image dataset, how can I 1) find this suitable prior distribution, 2) train your model on it?
I have some problem when the size of image h not equal to w, when cal psnr on phase B the render image is [w,h], but the real img is [h,w], this problem also show on tensorboard 'rgb'
detail:
similarity.py
mse
value = (image_pred - image_gt) ** 2