Lately I have been working on image classification problems but just recently I tried a new type of Problem on Kaggle.
I have normalized image data(X) by dividing with 255 and also the keypoints coordinates (y) by dividing with 96 as maximum height and width is 96.
So lastly model architecture in keras, I have tried smallest vanilla network spanning to Huge ones similar to VGG models and with epochs it doesn't seem to learn instead underfits or overfits.
And one other problem is dataset size is small by my best efforts I doubled the datset by flipping image and shifting values accordingly.
Even after that it seems to get nowhere. I think I'm done improving data but can you help me by giving me a model architecture that works.
Link to Kaggle Competition: https://www.kaggle.com/c/facial-keypoints-detection