hello everyone! I had to ask a question that i am fine tuning resnet50 to differentiate between flipped (mirrored) and non flipped images, i am freezing all its layers and adding additional layers on top but its doesn’t seem to give good results, i also used transfer learning method but it also gave 50% accuracy on validation set, I cant figure out how to train it better. can anyone have idea where I might be wrong?
What kind of images are you trying to classify? Is it easy to understand if they’re flipped (for example if they are text-based) or could they be easily confused at a quick glance (for example, graphs)
yes, so basically at first i was using completely random imgnet and coco dataset but as the accuracy wasn’t much increasing so i tried training it with the dataset of faces (human faces) only. I thought model could be able to differentiate between left or right sides of human features but yet no difference in results. But if you ask, flipped picture of, whether it is a face or an object, is very difficult for a human to differentiate. I’ve tried it myself. All in all images are quite confusing.
I believe it will be incredibly hard to accomplish this.
AFAIK most training datasets already flip, distort & rotate images to increase the training size, variability, and help understand images with incorrect orientation. I can understand and would use the same path as you, but I wonder if it’s fundamental to most, if not all CNN models.
I’m wondering if you would need to basically re-train from scratch to even begin the chance of having a usable model.
3 Likes
ok, Thanks for your help.