Need your help creating a masking tool!

Discussion in 'General Discussion' started by Centurione, Apr 14, 2017.

  1. Centurione

    Centurione New Member

    Thank you Iren! Got it ;)

    First of all thank you for all the masks I gathered till now - it's a huge help!
    Secondly - I need your help again! I was wondering whether you know any artists or studios that could share their anime-styled art or sample animation for me to use in the paper itself? While no one is going to look at the dataset I used, everyone will look at the paper and I would like to make sure all images used as samples to show results are 100% legal. I am pretty certain I can use Morvena project but it would be awesome if I could show results on different styles.

    And lastly, if you're interested in the topic, some time ago this repository

    https://github.com/matterport/Mask_RCNN

    showed up implementing a Mask R-CNN deep network based on this

    https://arxiv.org/abs/1703.06870

    paper and I've decided to use it as a backbone of my program. Still learning the network but it looks promising but not something of much use yet.

    So yup, still some work to do!
     
  2. Enigmo

    Enigmo Member

    Are Dragon Ball, Dragon Ball Z and Dragon Ball GT masks accepted, too? Or, are they too old for the type of animation you are looking for, Centurione? It's the only series I've ever masked, haha. :p
     
  3. Centurione

    Centurione New Member

    Sure it's alright. Any mask is cool. At this moment I use only anime-styled samples to improve the quality of model output and not distract it with completely different samples.
     
  4. Replay

    Replay Administrator Staff Member

  5. Centurione

    Centurione New Member

    If you have something to send, I'd be glad to accept it :)

    General update: I have finished my thesis and I am now officially an Engineer :D I need to do some editing so I will link the thesis in a couple of days.
    The results I got are, in my opinion, awesome considering the amount of data I have been working with. While I appreciate all the help, I did not manage to receive enough samples for the model to have a chance to do some real learning and it's basing mostly on the weights trained on MS COCO image set.
    Still, doing the research would not be possible at all without your help so I am more than thankful! I'm considering continuing the work as my Master's thesis.

    About the application itself: to deploy the interface quickly I made it as a web application with simple upload and download functions.

    And here are some bad news: I had a repository only for the web app and didn't store the final version of code for the learning process. I was doing the whole thing on my boyfriend's PC as it's the only machine I had access to that could manage such calculations in reasonable time. A few days before the defense he needed some storage and was cleaning some things up and kinda misclicked the starting letter of the folder and deleted it. And deleted means deleted, not moved to trash. While I have most of the code and restored crucial files for now I'm too tired of the project to spend my days trying to restore the whole configuration as it took us a couple of days to deploy the whole thing together. Though I will do it for sure because, as I mentioned, I plan on continuing the project.

    Here are the empirical examples of the best model in working:
    [​IMG]

    So, as you see, the results are faaaar from what you could use in production, but you can easily see that it generally works! I have some ideas for trying out models which are more segmentation-oriented or making a custom one. It also producs a low resolution mask as it would be a killer in regards of time if I fed it with higher resolution data.

    To sum up: there is potential but the thingy doesn't work well enough yet. Still well enough to get me a BSc though :D
     
    Kroner, Replay and CrackTheSky like this.
  6. Replay

    Replay Administrator Staff Member

    interesting results so far
     
  7. MagicDarkLight

    MagicDarkLight New Member

    To see it does not really follow straight lines with strong contrast is interesting, I would have expect it to do more straightforward masks. Let us keep posted about it!

    Congratzz also for becoming an engineer o/
     

Share This Page