0

We previously had 411 size data set. With that we got 70 % of the objects detected. We augmented it to 5000 by implementing rotations and different levels of brightnesses. But the accuracy dropped to 40%. We don't know where it went wrong.

Miki
  • 40,887
  • 13
  • 123
  • 202
Phaniraj Kattamuru
  • 1
  • 2
  • depends on whether your augmented data is more similar to the test data than your non-augmented training data... If your training and your test set for example only has "upright" persons, it doesn't make sense to rotate 45 degrees during augmentation, for example. – Micka Jan 16 '20 at 19:46
  • did you augment before or after the train/test split? – CAFEBABE Jan 16 '20 at 19:59

1 Answers1

0

It doesn't necessarily sound like anything went wrong. Instead, it sounds like your model can't handle rotations or varying brightness very well. If you expect to be dealing with those kinds of variations in your input data, you should probably increase the complexity of your model (or perhaps find a way to preprocess data to be more easily processed by the model).

Mark Snyder
  • 1,635
  • 3
  • 12