I’m starting to play with the Emotion SDK and setup the Video-demo (c++). I have it analyzing 10 second HD videos with clearly viewable faces (example: https://stock.adobe.com/stock-photo/businesspeople-listening-to-speaker-at-conference/55317473?prev_url=detail). I have the latest “data” folder for the SDK. The problem is that the accuracy/quality is quite poor. Meaning, in the referenced video the output has:
- Age ranges wrong - most are classified as “under 18” or “unknown”
- Ethnicity is either “caucasian” or “unknown”
- Gender is about 80% right, with a few “unknowns”
- Emotions are spotty - there’s a lot of zeros (or near zeros) in the data whereas, as a human, I interpret different emotions than the SDK.
Do I need to further train the system based on my videos?