Hi, thank you very much for your great work! Currently, I am using a modified diver model for one of my research. In this model, the size of the first layer is different (bigger), so I would like to know in detail how the pre-multiplication logic works, to see if I can premultiply my first layer with the features.

To the best of my knowledge, we keep the layer bias 0 "new_weight['weightin'][:,0]," and then perform the premultiplication to the features. Why we perform the transpose? "features = (features@new_weight['weightin'][:,1:].T).contiguous()", in this case, the shapes of the weights are 32x32, what happens if i change the size of my layers?

params = torch.cat([
                new_weight['weightin'][:,0],
                new_weight['weight0'].T.reshape(-1),
                new_weight['weight1'].reshape(-1),
                w2.T.reshape(-1),
                new_weight['weight3'].T.reshape(-1),
                torch.zeros(28),
],0).contiguous() 

# store feature vectors in an 1d array
voxels = new_weight['voxels'].coalesce()
N = voxels.shape[0]-1
idxs = voxels.indices()
features = voxels.values()

# pre-multiply first linear layer to feature vector for diver32 model
if voxel_dim == 32:
    features = (features@new_weight['weightin'][:,1:].T).contiguous()

Thanks :)

Hi! First of all, thank you very much for your fantastic work and code!
I am trying to optimize the c2w matrix(rotation and translation {R, t}). Now after loss.backward(), the gradient of R can be calculated, R can be optimized, but the gradient of t is calculated as None. Do you have any suggestions for this？

Hi!

First of all, congrats for your work! I really enjoyed reading the paper ang going through the code (which is quite understandable). I have a question about how you integrate the trilinear interpolation of the features for each voxel. In the paper, you obtain the following equation:

I do not see how you finally solve the integral and how you implemented. I undestand the implementation is the following one on integrator.py:

`

A = P.sum(0).prod(0)/12+(P.prod(1).sum(0))/6
X = F[1] - F[0]
X = X[0] - X[1]
X = X[0] - X[1]
feature = A[:,None]*X

P_roll = P[:,[1,2,0]]
# 3xBxM
B = (P*P_roll).sum(0)/3+(P*P_roll[[1,0]]).sum(0)/6
# 3xBxMxC
Y = F[0,0,0,None]-F[[0,0,1],[1,1,0],0]\
   -F[[0,1,0],0,[1,0,1]]+F[[0,1,1],[1,1,0],[1,0,1]]
feature = feature + torch.einsum('dk,dkc->kc',B,Y)

# 3xBxM
C = P.sum(0)/2
# 3xBxMxC
Z = F[[0,0,1],[0,1,0],[1,0,0]]-F[0,0,0,None]
feature = feature + torch.einsum('dk,dkc->kc',C,Z)
feature = feature + F[0,0,0]

`

What is the actual formula you are applying here? Do you have some derivation of the integral and this implementation?

Thanks in advance!

Hi @lwwu2,

First of all, thanks so much for the great work and for making the code available!

I was trying to train a DIVeR on the drums sequence from nerf-synthetic dataset. The readme for the training is unclear to me and I got some very weird results. Below is my way of training:

1. Train an implicit model in the coarse level (by replacing configs/config.py with configs/nerf_synthetic_coarse.py), and run python train.py --experiment_name=drums --gpus=1 --resume
2. Prune this trained coarse model via python prune.py --checkpoint_path=logs/drums/version_3/checkpoints
3. Continue training on the fine level (by replacing configs/config.py with configs/nerf_synthetic_fine.py), and run python train.py --experiment_name=drums --gpus=1 --resume.

Right after step 3, I obtain something very strange already, as shown here: https://imgur.com/a/mz6ILXq. I wonder which step I made a mistake? Also, could you please make the training step more clear so we know how to really reproduce the results shown in the paper?

Thanks so much in advance!

Best, Songyou

Hi, thank you very much for your great work! I can get the correct rgb image with your code, but the depth map is not very accurate. I am using the following code to render the depth map.

color, sigma, mask, ts = model(ray_o, ray_d)
depth = (ts*mask*weight).sum(1)

The rendered result(left) and ground truth(right) of the depth map is as follows. The rendered result is more like a grayscale image but not depth map, Do you have any suggestions for this?

Hi, congratulations on your paper. The results look great and I can't wait to try it on my own data! Would it be possible for you to add instructions as to what is required?