Improve Accuracy by Optimizing the Recognition Network
Face recognition uses a neural network to determine how closely a face in the ingested media matches the faces in your training database. The default neural network that is supplied with Media Server can be used with any face database.
To improve recognition accuracy, Media Server can fine-tune this neural network for the data in a single face database. To use the fine-tuned network, a face recognition task must use only that database, and the database must meet several requirements:
- The database must have been created with fine-tuning enabled (use the
finetuning
parameter in the actionCreateFaceDatabase
). - The database must contain a reasonable number of faces. OpenText recommends more than 50.
- The database must contain multiple training images for most or all of the faces.
NOTE: OpenText recommends that you fine-tune the neural network only if your training data is similar to the media that you will ingest. For example, if you are planning to ingest video captured from a CCTV camera, the training data should ideally have been captured from CCTV. In some cases, using action=FinetuneFaceDatabase
can result in overfitting and reduce recognition accuracy. If you find that accuracy is reduced, you can revert to the default neural network by setting DisableFineTuning=TRUE
in your face recognition task.
To optimize the neural network, run the action FinetuneFaceDatabase
. Run this action as the final step, after all other training is complete. OpenText recommends running the action BuildAllFaces
first, to ensure that all of the images have been trained. For example:
/action=BuildAllFaces&Database=politicians /action=FinetuneFaceDatabase&Database=politicians
The optimized network is used automatically, as long as you set the Database
parameter in your face recognition task. For more information about configuring a face recognition task, see Recognize Faces.
After optimizing the network, you can add faces to the database, or remove them, as normal. However, OpenText recommends that you run FinetuneFaceDatabase
again whenever you make significant changes to the training.