i resized to 224x224 with some tweaks managed to get 60% val accuracy.I used some image processing techniques but doing that reduced my val_accuracy
Key thing in Image augmentation is to choose what to use for what type of dataset and problem statement. Think about each augmentation whether that is a realistically possible or not, even with the value of augmentation.
For example think you have cars dataset, if you add vertical flip, do you think it’ll be helpful since you’ll be training to a image which is not possible or a very rare case.
Also visualizing the augmentations that you apply helps.
If you are interested to read further regarding better augmentation policies , you can go through paper (https://arxiv.org/abs/1805.09501), where you can find a way to auto-find effective augmentation policy in given search space.