Facing issue with High resolution images
As we do inference on High resolution images and after applying the threshold to make the mask binary mask with solid black and white that time it is creating some uncertain repeated pattern which decreasing the accuracy of the results
second issue if we not apply threshold to mask the solid black and white mask then it is creating some grey pixel inside the foreground which is also not acceptable for the good accuracy of the results
Hey, thanks for reaching out can you share the input image for testing?
You can test on this images but use thresholding to create the solid black and white mask and let me know the results
Pattern like this there should be straight line but it is crating some repeated pattern which is decreasing the accuracy of the mask
hello
Do you have any solution for this .......... let me know if there is any solution to avoid uncertain staircase in the mask
Hello can you please reply if you have any idea How to deal with this problem ............ as this is the very serious issue and your model will also face this issue which decreasing the accuracy of the segmentated mask
@schirrmacher hey is it possible to connect with you ...... i also have some expertise on this model and also solved lot of issue may.. can you have some discussion
Sorry for the late response I am currently traveling… What is the size of your original image? Images are scaled down to 1024x1024 pixels for creating a segmentation mask. I can imagine that a big input file might scale this mask so that it has such defects…
Thanks for the response .....yes it is because of upscaling then input image from 1024 to its original resolution .. do you have any idea how to deal with this issue because my most of the work contain high resolution images only and i am stuck in this problem from long time ........... also currently i am working on to reduce the blur effects on the edges
I once had the idea to upscale the input and output architecture. But before doing that I want to improve the quality of the current status and cover edge cases. Additionally I have no clue if a bigger input/output network can be easily trained or if training costs would explode…
Can you tell me what the size of the images is you used? Or is it a typical camera image, e.g. 4032x3024 pixels?
i was using the image of 4320 * 6000 resolution image ........... i also trained the model on 2048 but it also has some limitation of the resolution that's why i am trying to solve this issue preventively also i tried with the patch based approach but problem with that was if their is o.1 error near edges then that stitches are visible .......... if you have any idea please suggest me i will implement it and try it
An other idea could be training an image upscaler for segmentation masks 🤔
Good idea to use image upscaler .......but my usecase is i don't want to use any another model except segmentation model
Have you trained the 2048x2048 input model with my dataset? Because the image sizes are not big enough. I am currently working on a dataset with higher resolution and more hair details…
No i did'nt train it on your dataset but i trained on my dataset with 2048 2048 resolution on high resolution images but their i used patched based approach ...................also when we infer model is generating the grey pixel at the blur part so in cutout that blur part is getting minus from the image ..............and we call it as inaccuracy ..now i am working on it where i want the solid mask with the threshold with the hair details with the transparency and also smooth edges
Hey will it be possible to connect with you .......may be will discuss our problems about the model ...... i was also working from long time on this model .....
We can discuss what you tried and potential ideas for solutions. For me its fine to do it here on hugging face or discord. What do you have in mind?
I created a discord channel: https://discord.gg/3PaqGFza
I have accepted your invitation on discord .....lets discuss thing on discord