Sir your videos and explanations everything is just awesome, it’s like gold mine for young grad researcher , May Almighty bless you and your channel deserves millions subscribers ❤️
10:15 the images enter and exiting SRGAN has the same resolution. For training first HR image is blurred to give a Low resolution image and it is downsampled meaning the resolution is increased again by a quick method like bicubic interpolation to make the LR training set. It is simply a resized blurred image
Hi Sreeni. I just need desperately for making a picture more clear. Do you know of anyone who can help? Your response and help will be greatly appreciated. Thank you so much.
thank you so much. But how can we train this ? I am facing low Ram and GPU issue even with paid google colab pro. how can we train these model please guide
Thank you for the amazing information! Question... for pixel classification let's say we use random forest and support vector machine, is there a way to accurately compare the results and not just visually?
Hey Sreeni, I really like your GAN tutorials. I was wondering can we use GAN to fuse multi modal images? Like the tutorial 251 where you used satellite images to map translation. What if we have more than one input i.e two satellite images lets say with different resolutions and we want to fuse those two images and translate it to maps. Like we have two Real images Real A and Real B (both satellite images) and only one map images(ground truth). How can we solve this problem? I would really appreciate it if you give a direction on this task. Many thanks and keep up the great work.
Two images and multiple resolutions mean they have different size pixel arrays. I don't know how you'd fuse them to get a meaningful representation of the field of view. If you have two images of different modalities, say visible and infrared images, then you can combine them into multiple channels and work with the multichannel image.
Hello sreeni. Can you please make video on adversarial training. Where we can add discriminator loss and generator loss. Like total loss = lambda*adv loss + generator loss It is useful in domain adaptation in segmentation. I'm not sure how we can add loss in Keats since we directly use train on batch. After that how can weodify the loss
SRGAN has already used adversarial training. You can refer to this video. If you use "total loss = lambda*adv loss + generator loss", the adv loss is GAN loss and the generator loss is the content loss (e.g., MSE). “adversarial loss" is captured by the discriminator and is not mysterious as the name seems to be.
U net doesn't add skip connections in the same way. The skip connections in this GAN are added to deeper layers, whereas the skip connections in the Unet's encoder are *concatenated* with the corresponding layers in the decoder