Accuracy of Video Indexer Person recognition based on the Face API

There is an issue with the Azure Video Indexer person recognition. My company uses the Video Indexer to recognize persons based on the person model. We noticed that if we divide the video into individual frames and use the Face API directly with these frames, the recognition rate for the persons is much higher than if we search for persons with the Video Indexer. That doesn't seem to be right. How does that happen? How exactly does the Video Indexer use the Face API? Which frames are generated? Is the video divided into frames every second? Or is the analysis based solely on the keyframe? As far as I know the Face API works with pictures only. I sincerely hope that you can help me. Thanks for your time!


    Hi, Video Indexer uses Face API for face recognition but not for face detection. The face detection model that Video Indexer uses looks at every frame but is optimized for videos (using the face API detection on a video will be much slower and costly) Would be good to add a user voice from the customer as well as to ask them for example videos for us to investigate if possible. Also, can you please elaborate on the customer scenario here? Do they have the face and want to identify what frame it is in? something else?
    Posted by Hidden Thu, 01 Aug 2019 12:41:13 GMT

You're not signed in. Please sign-in to report an issue or post a comment.