I can't describe how much your "U-Net Semantic Segmentation" helped me! They are all super super superrr helpful! Thank you for sharing your knowledge in this way. So clean, clear, organized and so so so ..! Thanks again!
Sir, you are doing a great job guiding us with such highly valuable guidance. I need help quickly, how can we use this technique you explained in this video to classify large histology images? Not segmentation but classification? Please help
A video about crop patches compared with random patches should be interesting. Since random patches does not crop evenly, instead it randomly crop with multiple instances, meaning each pixel have a chance of being present in multiple crops, then using Voting to classify the pixel. Random crop is interesting because some pixels in the edges of each crop might require further information "context" of it's surrounding pixels in-order to classify itself, thus having multiple random crops ensures that the prediction of each pixel is based on the vote of multiple random patches with different surroundings "context".
At the end you want a segmentation result of the large image. So you need to segment every pixel in the image anyway. I do not see any advantage of random crop during prediction. It does make sense during training as it helps generalize the model better. The smooth blending operation, used in this video, performs image transformations and segments the transformed patches, then finally blends to average all the predictions. In a way, this is similar to your proposal. Your thinking is in the right direction, thanks for the comment.
Just a hint regarding matplotlib colors. Add vmin and vmax param where vmin is the min class integer and vmax is the max class integer of your segmentation mask. This will keep the colors constant.
Thanks for alerting me about the missing file. It has been uploaded. Here is the direct link: github.com/bnsreenu/python_for_image_processing_APEER/blob/master/smooth_tiled_predictions.py
Great teachings and codes 👍🏽, need help with this: IS IT TOTALLY OKAY TO RESIZE VERY LARGE IMAGES SAY 10000×10000 TO 128×128...YOUR RESPONSE WILL BE HIGHLY APPRECIATED.
outstanding work sir........I got this error NonUniformStepSizeError: Unpatchify only supports reconstructing image with a uniform step size for all patches. However, reconstructing 17 x 256px patches to an 4408 image requires 259.5 as step size, which is not an integer. Please help
![Noob To Max With DRAGON REWORK In Blox Fruits [FULL MOVIE]](http://i.ytimg.com/vi/LnBMOinoOvA/mqdefault.jpg)
I can't describe how much your "U-Net Semantic Segmentation" helped me! They are all super super superrr helpful! Thank you for sharing your knowledge in this way. So clean, clear, organized and so so so ..!
Thanks again!
Sir, you are doing a great job guiding us with such highly valuable guidance. I need help quickly, how can we use this technique you explained in this video to classify large histology images? Not segmentation but classification? Please help
you effort is highly appreciated .. thanks a lot
It's my pleasure
A video about crop patches compared with random patches should be interesting. Since random patches does not crop evenly, instead it randomly crop with multiple instances, meaning each pixel have a chance of being present in multiple crops, then using Voting to classify the pixel. Random crop is interesting because some pixels in the edges of each crop might require further information "context" of it's surrounding pixels in-order to classify itself, thus having multiple random crops ensures that the prediction of each pixel is based on the vote of multiple random patches with different surroundings "context".
At the end you want a segmentation result of the large image. So you need to segment every pixel in the image anyway. I do not see any advantage of random crop during prediction. It does make sense during training as it helps generalize the model better. The smooth blending operation, used in this video, performs image transformations and segments the transformed patches, then finally blends to average all the predictions. In a way, this is similar to your proposal. Your thinking is in the right direction, thanks for the comment.
@@ZEISS_arivis Thanks
Just a hint regarding matplotlib colors. Add vmin and vmax param where vmin is the min class integer and vmax is the max class integer of your segmentation mask. This will keep the colors constant.
Thanks for the tip
I tried to make this run using pytorch tensors but it didn't work
Great video. But I am bit confused about overlap tile strategy in Unet paper. Can anyone give me some clarity over it
How do i work with this 564x584 image size
How do i divide this image into patches when tje height and width of the image are not equal
Thank you very much.
Great job! Thanks!
Glad you liked it!
Thank you so much for your efforts 😍😍
My pleasure 😊
the file "smooth_tiled_predictions.py" does not exist in the repo.
Thanks for alerting me about the missing file. It has been uploaded. Here is the direct link: github.com/bnsreenu/python_for_image_processing_APEER/blob/master/smooth_tiled_predictions.py
Great teachings and codes 👍🏽, need help with this: IS IT TOTALLY OKAY TO RESIZE VERY LARGE IMAGES SAY 10000×10000 TO 128×128...YOUR RESPONSE WILL BE HIGHLY APPRECIATED.
For classification purposes?
outstanding work sir........I got this error
NonUniformStepSizeError: Unpatchify only supports reconstructing image with a uniform step size for all patches. However, reconstructing 17 x 256px patches to an 4408 image requires 259.5 as step size, which is not an integer.
Please help
I guess the error is clear enough where it says you need to define an output size that is a result of integer step size.