Evaluation metrics: LPIPS,SSIM,PSNR

[xiaoyudanaa]

Many congratulations on such an excellent article, I had some problems reproducing the results. I have experimented with the training weights you posted on the jumping dataset, but the final result is very poor as shown in the figure below, I would like to ask what causes this? I look forward to your reply. (The original set chunk size=8192, but the GPU could not run due to out-of-memory problem, so I set the chunk size to the same 1024 as NSFF)

[zhengqili]

Did you correctly load the model? This is usually due to the fact that you point to the incorrect model path.

[xiaoyudanaa]

Based on your response, I checked my configuration path and its doesn't seem to be wrong. The details are shown in the attached image：

[zhengqili]

When you run the script, from the terminal did you see something like "Reloading from xxxx, starting at step=yyyy" or "No ckpts found, training from scratch..."?

[adamInThe80s]

Hi, I'm not the author of the paper. But I ran into a similar issue as you.
In ibrnet/model.py line 104:
out_folder = os.path.join(args.rootdir, 'checkpoints/fine', args.expname)

The "fine" model checkpoint location is at a hard-coded location relative to the root directory.
So if you are correctly pointing to the coarse checkpoint location in the config file, you might still be running to this problem.

[quan5609]

Hello @xiaoyudanaa, do you have any summary of per-scene results on nvidia dataset?