Depth bug when importing YML

[kaelsonofkrypto]

i finally was able to test importing my own depth maps in yml, i deflickered the iw3 depth maps in davinci and then i replaced the files in the export disparity "depth" folder with my new deflickered maps, and imported the yml file in iw3. So what i found out is depth 3.0 was the max u can push before the image begins aliasing/jagged edges. i also noticed the depth looked no different on 5.0 depth but significatly looked worse with jagged edges, ironically became lower resolution, and got flatter in background when going above 3.0 on depth, so i had a hunch, i lowered depth to 2.5 and image became suddenly more 3d-ish and clearer, then when i lowered to 1.8 depth, image became even more 3d-ish. so the depth function doesnt not work properly when using YML method. So on yml method its better to use a low depth of 1.8. essentially 1.8 depth on yml method looks equivalent to 3.0 depth on the regular 2d to 3d method. The regular 2d to 3d function doesnt have this bug , it works fine if u just do a regular 2d to 3d conversion with no yml, but the moment u use yml method u actually have to use less depth for it to function properly

[diverswan9]

I've been experimenting with .yml as well. Just to see, I loaded a bunch of different models and they all came out pretty much the same. It must be mainly using the depthmaps for what it needs.

[nagadomi]

If there is no problem importing without editing after export, then it is a problem of editing.

One cause that I can think of is the resolution (image size) of depth images.
The row_flow_v3/v2 model is performed on the depth input resolution. It produces a larger stretch the larger the image.
If you have resized the depth image for editing, you will need to resize it back to the original size to get the same result.

[kaelsonofkrypto]

thanks for the feed, this is good information that would be beneficial i believe in the readme section where u have info on different sections of the program.

There are a few questions i had that i couldn't find info on that feed.

1.For the depth resolution setting, depth level, and the converge, do those 3 get applied on export disparity to depth map or do they get applied on import of yml?

2. i know u mentioned that row flow is trained on 1920, so then if i use a 4k or 8k source, then should i be selecting an output limit of 1080p? is this why there isn't an output limit of 4k option available in that output limit section?

3. also does tta get applied during export disparity or only on import, and does it make more sense to use fp16 on export disparity and tta on import yml? trying to ensure the depth map and conversion is accurate as possible but if its not working on a certain phase i want to make sure i apply tta to the correct phase. i do like tta as i can tell the different, it has less flickering than fp16

4. also on the feed i read you mentioned edge fix (edge dilation) and mapper (foreground) get applied on depth which i assume you mean this gets applied to the depth map when you do export disparity, and it does not get applied on video conversion which i assume is basically the process of you importing the yml file, and also flicker reduction (Ema normalize) gets applied on both export disparity of depth map and on import of yml for video generation. Did i understand this correctly?

and please forgive me for all these questions, im just trying to make sure i get this right.

[nagadomi]

2

I added the Stereo Processing Width (--stereo-width) option.

This allows you to specify the size (width) of the depth image to be input to row_flow_v3/v2. Internally, the depth image is resized to the specified size before processing.
For example, if importing 2K depth images, you can specify 1280 or 1920 to be processed as a safe size. By specifying this option, even if the depth image is 2k to be imported, 2 is not needed.
Also, it can be used when you want to run row_flow_v3 at a higher resolution than Depth Resolution. (Not tested in detail, but may be able to reduce artifacts)

1

Depth Resolution is used only when Export, and 3D Strength and Convergence Plane are only used when Import (yml input).

3

TTA is only used when Export.

Basically, Export/Export disparity only outputs depth, so the stereo generation related settings are not used.
The confusing parts are Foreground Scale and Edge Fix. They are applied at different times in Export and Export disparity. However, they are only applied when either Exporting or Importing.

[kaelsonofkrypto]

Thanks so much, appreciate the addition and the explanation. Can't wait to try it out.

[diverswan9]

Is that update supposed to be for export/export disparity only? Can it be used for a normal 2D-3D conversion? I tried applying different settings for 'Stereo Processing Width' on a normal conversion (not an export) and didn't work so well. I

[nagadomi]

The reason for adding Stereo Processing Width is for Import (yml input, solution for cases where the depth image is too large).
It can be used for other than Import, but the effect is different in the case of downscaling and upscaling.
For non-Import uses, it is (bilinear) upscaling in most cases, it may not make much difference, or it may be worse.
If you do not feel it makes sense to use it, I recommend Default.
Default means do nothing. It is the same behavior as before.