We present a novel online unsupervised method for face identity learning fromvideo streams. The method exploits deep face descriptors together with a memorybased learning mechanism that takes advantage of the temporal coherence ofvisual data. Specifically, we introduce a discriminative feature matchingsolution based on Reverse Nearest Neighbour and a feature forgetting strategythat detect redundant features and discard them appropriately while timeprogresses. It is shown that the proposed learning procedure is asymptoticallystable and can be effectively used in relevant applications like multiple faceidentification and tracking from unconstrained video streams. Experimentalresults show that the proposed method achieves comparable results in the taskof multiple face tracking and better performance in face identification withoffline approaches exploiting future information. Code will be publiclyavailable.
展开▼
