Since I am getting a low accuracy(around 50%), and after using the cleaner, I start to think of some factors that could possibly cause high error rate. Here is my thoughts:
-
Just like why we want to use the cleaner, the google image result might contains images that do not belong to any target categories.
-
The image, pixel compressed or not, has no clear features for the model to learn. For example, I try to train 12 shoe brands of in total ~1000 images. The fact is some shoe images have very faint logo(you can barely tell the brand even with naked eye except that it’s a shoe). The model could possibly not be able to train them.
-
Even for different brands of shoes, they can have the same design, color and size except foe just the logo. The model might be confused in this case.
Can anyone think of other possibilities or correct me if I was wrong.