Before you start
- Camera frames are sent to this server only to compute a prediction. They are held in memory for a few seconds and are not stored or logged.
- The model labels visible facial movement. It does not read feelings, honesty, or intent.
- Accuracy is limited: about 63 % average recall on the DFEW benchmark without test-set tuning, lower for Fear and Disgust. Do not use it for decisions about people.
Live prediction
Starting camera…
Upload a short video
One face, a few seconds, at most 30 MB. The file is deleted right after the prediction.