Is this net a NeRF model for this line of code in the pose_estimation.ipynb file?
# Encode. net.encode( input_image.unsqueeze(0), input_pose.unsqueeze(0).to(device=device), focal, )

Are the parameters of this NeRF model modified during iteration?

On the other hand, the net.encode seems does the same thing in each loop, but why can't move this code outside the loop?

Hi, thank you for your excellent job,
I'm pretty new here, and I have a question.
Usually, each NeRF model saves the network as checkpoints (.ckpt)
How come your trained model in "iNeRF/pixel-nerf/checkpoints/ is some binary files? My goal here is to use other trained networks in checkpoint format.
Thanks in advance

Hi， thanks for your great work!
May I know how your NeRF models were generated? Were they generated by the official NeRF code? Or were they generated by thepixelNeRF as mentioned in the original paper?

Regards,
William

When I compare the optimized camera pose and the real relative pose transformation of the two input images, there is a large deviation, especially when calculating the rotation angle of phi. May I ask what transformation should be done before comparing? I don't know which part is the problem
Below is the part of the code I compared:

optimized relative pose transformation

pose_pred = predicted_poses[-1].copy()
pose_relative = np.linalg.inv(input_pose) @ pose_pred
print('Pose prediction:', np.arctan2(pose_relative[1, 0], pose_relative[0, 0]) * 180 / np.pi)
Output: Pose prediction: 2.1300169042266903

the real relative pose transformation

pose_convert = np.linalg.inv(input_pose_np) @ target_pose_np
phi_convert = np.arctan2(pose_convert[1, 0], pose_convert[0, 0]) * 180 / np.pi
print(f"Phi int: {phi_int}")
print(f"Phi target: {phi_target}")
print(f"Phi convert: {phi_convert}")
Output: Phi int: 155.91559686333247. Phi target: 119.24629196439332. Phi convert: -1.5495584242778404e-05

Hi, thanks for your code!
It seems that the pose optimization in this code is different from that in the paper of iNeRF?

Hi Authors,

Thanks for the work!

I tried to use the provided code on some new dataset. However, within this code base, I only find three sampling method, namely "random", "center" and "patch" at here. Correct me if I'm wrong, but I don't think any of them correspond to the "Interest Point Sampling" or "Interest Region Sampling" mentioned in the paper which are shown to improve the performance.

Could you point me to the code for these two interest sampling methods if they are in the code base? Otherwise, will you release the code for them?

Thanks!

Thanks for your excellent work!But can iNeRF work with Instant ngp,which is faster and better that the original NeRF?

Hello @yenchenlin,

Thank you for sharing the repo for this great work. I enjoy reading your paper...
By the way, the teaser image and demos look fantastic!

Here, I am currently trying to run your code, but my GPU RAM is only 9 GB.

So, I would like to kindly ask if there is any way to run your code under this limited GPU RAM situation....?

Hi, i try to run the code on the colab, and I find that the code run perfectly on my local laptop(fail because of fun out of PU memory), but on the Colab it throw out an error call "parseexception: expected '}', found '=' (at char 759), (line:34, col:18)" in cell "def extra_args(parser):". It is quiet weird. Because I think parser arg is greatly written, on the other hand when I run the cell locally there is no problem at all.
Did u have any idea?