Based on the latest version, I got this error: FileNotFoundError: [Errno 2] No such file or directory: './data/DTU_MVS//split.pkl'

I tried to copy the original split_shuffle.pkl to that location and it works. I guess you should not remove that file?

Thanks for this valuable work!

I failed to run pip install torchsearchsorted/. I have Ubuntu 18.04 + GCC v8.4 + CUDA 10.1 + python 3.6.12 + pytorch 1.6. But on another Ubuntu 16.04 (GCC 5.4) machine, it can be compiled successfully.

Is that an error due to any package version issues? Could you please help? Thanks!

Here I show one of the error messages (more detailed error messages can be provided upon request):

    ninja: build stopped: subcommand failed.       
  Traceback (most recent call last):                    
    File "/home/t-yizhouwang/anaconda3/envs/srf/lib/python3.6/site-packages/torch/utils/cpp_extension.py", line 1515, in _run_ninja_build 
      env=env)                                                                    
    File "/home/t-yizhouwang/anaconda3/envs/srf/lib/python3.6/subprocess.py", line 438, in run
      output=stdout, stderr=stderr)  
  subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1. 

Hi, I am attempting for novel view synthesis by running the pretrained inference with the provided command and code. Even if I reduce the number of ray samples, the memory was exploding (500G memory + a lot of swap on the machine). Does anybody else experience this? or could the author kindly comment on this?

Hi, I am using a Tesla V100, and I can only run the training script with N_rand set to 64, but the default N_rand is 4096. Are you using that big batch size during the training?

Hi,

Thank you for releasing the code.

I am trying to train this model on my dataset. I just revised the dataset-setup code. The resolution of my image is 2000x2000. and I set ref_imgs number to 5. The GPU is RTX 2080 with 12GB memory. I try to train this model several times, but I got the following err message:

So I want to ask the requirement of GPU memory for training this model.

Best, Hengfei

Hello jchibane, I'm wondering whether your code can be applied to other datasets such as ShapeNet and custom blender datasets. How did you get the raw pose_extrinsics for the DTU dataset ? Thanks.

When I run inference using the latest code, I got this error: KeyError: 'sim_combinations' at https://github.com/jchibane/srf/blob/ce97b2797578a96818c2b53b3a38d34218d2f55f/model.py#L213

I guess the pre-trained ckpt files are out of date without this key included? (Comment this line out will work for inference.) Thanks!

Hi, I was hoping to obtain the 3D mesh as shown in Figure 6. I assume you'd follow the steps to threshold the density predictions and run Marching cubes on a dense grid of points. Could you perhaps provide the exact detail? or upload the code to obtain the paper's Figure6 mesh output?

thanks,

I want to try to finetune on new scenes but it is too slow to download all the DTU datasets. So I want to finetune SRF directly based on the pretrained model on DTU. Could you share the pretrained model on DTU before finetuning on Scan106 or Scan23?

Dear author,

can you give more explainations to following two coordinate system? x_down_y_down_z_cam_dir x_down_y_up_z_neg_cam_dir

How can I draw to visualize the xyz axises?

Thanks

File "E:\OTHERS_WORKS\srf\srf\dataloader.py", line 101, in getitem imgs, poses, poses_idx = DTU.load_scan_data(sample, self.load_mode, self.num_reference_views + 1, self.cfg, self.load_specific_reference_poses, self.load_fixed, self.shuffle) File "E:\OTHERS_WORKS\srf\srf\data\load_DTU.py", line 59, in load_scan_data poses_idx = random.sample(images[scan],num_views) NameError: name 'images' is not defined

Hello, I was wondering how the near and far values were decided upon for ray sampling. The near value = 500, far =1000 at the moment, how were this decided? and if I were to be trying on a different dataset, on what basis should I choose and near and far values?