why are there weird red margins of my loaded images?

[elenacliu]

Describe the bug
When I load the images, they have red margins.

To Reproduce
Steps to reproduce the behavior:
I just install the latest version of nerfstudio (commitid 9b6010e), and pip install -e . to install it.

Expected behavior
A clear and concise description of what you expected to happen.

Screenshots
If applicable, add screenshots to help explain your problem.

Additional context
Add any other context about the problem here.

[brentyi]

Hm, I'm seeing something similar:

[tancik]

I believe those are the poses used for eval.

[elenacliu]

I use the instant-ngp data parser, so all of the images are used for eval by default?
When I use nerfstudio data parser, only a few images have red margins.

[tancik]

@FrederikWarburg any idea what's going on?

[FrederikWarburg]

Hi!

I had a look at this. It seems like InstantNGP(DataParser) is using all frames for training and for eval. def _generate_dataparser_outputs(self, split="train"): takes split as input, but the variable is unused. Therefore, the viewer first plots all the training images without a border (aka all images), and then, on top of the training images, plots the eval images with a red border (aka also all images).

So, I guess we could

1. add train/eval split to the InstantNGP(DataParser)
2. maybe remove InstantNGP(DataParser) as it seems to be almost the same as Nerfstudio(DataParser). I guess the main difference is get_focal_lengths, but could easily be added to Nerfstudio(DataParser)

In short, I don't think the problem is the recent commit, but rather that using InstantNGP(DataParser) reuses all training as eval images. Wouldn't mind doing any of these changes, so let me know what you think is the best way forward?

[elenacliu]

@FrederikWarburg thank you for your reply. I may prefer the first way, because many of my scripts just use instant-ngp-data, and I don't need so many complex preprocess in other dataparsers.

upd: I have fixed the issue in pr #2335