Thursday, June 4, 2020

Winslow Homer's "The Adirondack Guide" - 2d to 3d image conversion using the3dconverter2

This post presents a 2d to 3d image conversion of Winslow Homer's "The Adirondack Guide". I undertook this tedious process mostly in order to make a video tutorial showing the whole process from beginning to end. As you can see/hear in the video, it's not that easy to make 2d to 3d conversions.


Reference image.


Dense depth map generated by the3dconverter2.

It took me about 2 hours of trial and error and I must say that I was completely drained when I decided it was time to quit and call it a day. This depth map has obvious errors which I am just discovering now. I completely forgot about the hole around where the left oar holder is. Also, I put the hole between the left arm and leg of the guide at the wrong depth. Last but not least, I completely ignored the depth discontinuity on the right side of the boat leaving a gigantic hole, which would prevent any boat from floating, I think. Those are easy fixes though.

The so-called Gimp paths are used to provide dynamic depth clues to the3dconverter2, as opposed to the static depth clues given in the sparse depth. I call the clues given in the Gimp paths dynamic because they can be easily changed either by modifying the path itself or by changing the path's name and hence its associated depth clue. Now, what does the3dconverter2 does with those Gimp paths? For the3dconverter2, any pixel along the path is assigned the depth given to the path itself (via its name). For any pixel with unknown depth (not along a Gimp path or not drawn in the sparse depth map), the3dconverter2 will give the best possible depth value by propagating the depth of pixels which are at known depths.

3d wiggle/wobble made with wigglemaker.

Video showing the whole process:


Follow me on youtube if you want more in-depth videos since those are a bit painful to make.

UPDATE!

After a lot of rest, I made a few changes to my Gimp paths here and there, especially for the boat. It really didn't take long, maybe 10 minutes, until I was satisfied. What really helped is to order the paths (in the path dialog window) w/r to depth.


New dense depth map produced by the3dconverter2.


New wobble/wiggle 3d animated gif.

Video showing the fixes:


Follow me on youtube if you want more in-depth videos since those are a bit painful to make.

8 comments:

  1. I like that you are going more in depth on how to use the software. I am very interested in 2d to 3d conversion for lenticular images an 2d to 3d conversion

    ReplyDelete
    Replies
    1. i try to but i always forget something. not easy to make tutorial videos. and time consuming too. Thanks for the comment.

      Delete
  2. I am glad you are going more into depth with the software. I am interested in 2d to 3d conversion and lenticular pictures

    ReplyDelete
  3. I have been following you for years trying to figure out how to use your stuff to create lenticular pictures from 2d pictures. These last few tutorials have been good in your explanation of how to build the files to make the software work. I know that using a slide bar and taking the views at one time is easier but the conversion steps are imperative when you are doing fine art photography and you slow shutter speeds down to capture water. I spent time in the Columbia River Gorge, in Oregon about 10 years ago now and noticed to create a 3d picture from a waterfall using 2 pictures wont work. If with some experimentation I can become successful with this, I would take some screen capture videos and help show this process off. I can see that 2d to 3d conversion can be exhausting but I think you could benefit from other readers and followers documenting their work and guest posting to help you out. Your blog has captured my imagination as far as 2d to 3d conversion. Now that you are spending time in the software and going over a step by step process, I think you should challenge your readership to use what you are showing and see if any would be interested in sharing. you being the owner of this would have to review and approve first. My overall goal right now is to perfect this so I can start converting some public domain pictures and start posting them on Facebook. I love that Facebook is developing that platform through there ownership of Occulus but there are some bad examples of 3d on there. There is a lot of room for improvement and I think your software can help.

    ReplyDelete
    Replies
    1. Thanks Larry for your comment. Well, I am always open to readers sending me their work an making a post about it. I have done it before. I would also welcome readers making their own videos showing how they use the software. I am so used to using my own software that I often completely forget to explain something that is not obvious. As you know, 2d to 3d image conversion is not a walk in the park (it depends on the piture of course) and so not too many people are willing to spend the time to do it right.

      Delete
  4. I am glad that you are publishing videos on yt. What I commented right away. Double joy that there are machine learning films.This field is evolving rapidly and is using depth mapping for a variety of purposeshere new and surprising opportunities for development emerge thats simple https://youtu.be/MrIbQ0pIFOg https://youtu.be/BjkgyKEQbSM

    sorry for mistakes i use google translator
    greetings

    ReplyDelete
  5. good-sized maps can be created with a notebook Google colab MiDaSv2 https://colab.research.google.com/github/pytorch/pytorch.github.io/blob/master/assets/hub/intelisl_midas_v2.ipynb

    ReplyDelete
    Replies
    1. Cool! I will check it out and probably make a video about it at some point.

      Delete