Ask Your Question

Revision history [back]

Your problem can be two-fold

  1. It is the -w 10 -h 10 object size setting. This only gives you 100 pixels to generally describe your object class. It might be possible that small motion blurs simply remove to much content.
  2. There was no blurring in your training data, and thus, your classifier overfitted to clean and clear data. To avoid that, you will need to add slightly blurred training samples to your training set.