Very poor prediction results with my images

Hi I have just started lesson 1 and swapped out the cat / dog images with images of my two kids which I thought would be rather fun, but the results are very disappointing – I chose 100 images, put 80 in train and 20 in valid for each of my two boys and results were all over the place, best true positive (TP) result was .31 for one boy and .79 for another, with worst false positive (FP) result being .81 for one boy and .26 for the other. Does anyone have thoughts on whether the following are contributing factors?

  1. Source image size (my files are around 500 to 4MB each)
  2. Rotation: I read that sometimes images are rotated and have EXIF instructions to orient this. ChromeOS and MacOS are correctly interpreting this but when I show them in Jupyter, some of the images are still rotated 90 degrees which I can imagine might affect things.

Have you tried doing the same thing, but for two different unrelated people? My guess would be that you children possibly look quite alike, so it is harder for model to differ than cats and dogs or different dog breeds. Maybe try lower learning rate for the last layers?

1 Like

Yeah great point :slight_smile: I will try. Any thoughts on whether the rotation-in-EXIF affects things also most welcome.