compvis / geometry-free-view-synthesis Goto Github PK

During the collab demo running, I stuck with this problem.
Can I do something with it?

Downloading: "https://github.com/intel-isl/MiDaS/archive/master.zip" to /root/.cache/torch/hub/master.zip

---------------------------------------------------------------------------

ModuleNotFoundError                       Traceback (most recent call last)

<ipython-input-8-377ddaf3164f> in <module>()
      9     print("Warning: Running on CPU---sampling might take a while...")
     10     device = torch.device("cpu")
---> 11 midas = Midas().eval().to(device)
     12 renderer = Renderer(model=model, device=device)

9 frames

/root/.cache/torch/hub/intel-isl_MiDaS_master/midas/vit.py in <module>()
      1 import torch
      2 import torch.nn as nn
----> 3 import timm
      4 import types
      5 import math

ModuleNotFoundError: No module named 'timm'

Hello,

Thanks for open-sourcing! Do you mind adding training/evaluation scripts for RealEstate & ACID?

I notice that in the multiembedder, the inverse intrinsic matrix is also included. Since this information should be already determined by K, why is this also used? Did you notice some improvement?

Thanks

Is this possible to run in colab with the demo? Infinite Nature was able to create a demo in Colab :)

I am trying to understand the code and had one query about the dataloading in acid and real estate 10k dataset. I will try explaining it using the acid dataset case. The label assignment for large/small, forward/backward movement is done before actually sampling the source and destination frame ids. According to my understanding, this should be done after the frame id selection is done. Can you please let me know if I am missing something?

is inference on cpu possible?

Hi @pesser ,
What is the minimum GPU VRAM required for training on the RealEstate10K dataset?

when using colmap to generate the new data.

create empty data. But in the training code, there are many parts need to process this empty data. Why?

And could I do not use colmap to generate new data and train this model? I do not get the point to use it because we already have intrinsic params and camera pose. And I do not find any part in code using colmap data like database data.

Thank you!

when I try to execute pip install geometry-free-view-synthesis, I got this error. Is that problem with the pip source?

I've used the "alternative" install method. I'm running into this when hitting SPACE in the scene.

Hello,

I am curious if the implementation would be capable of novel view synthesis of objects if trained on multiview object data, as opposed to scene data? Do you see any potential blockers?

Thanks!

How can I include multiple input images into the synth? When starting braindance.py path-to-my-imagefolder, a filemanager opens that just let's me select a single image.

I'd like to re-create a whole room by including the info's of multiple images.

Hi,

Could you share the Rel10K training images for the first stage training?
We would like to train a different first stage model with the identical training data for fair comparison.

Best,

hello,

when i run pip install geometry-free-view-synthesis, proper version of splatting seems missing,


Any suggestion about this problem?

Hi, I downloaded the RealEstate10K dataset, and it only contains the txt files.
Can you attach your code for downloading the dataset from those txt files?
Thanks!