Abstract
In this article, we report a study on the problem of person identification in TV programs, such as situation comedy shows. A person identification system is constructed based on the joint use of visual and audio information. The system consists of two modules, namely, the analysis and the fusion modules. The analysis module contains a visual analysis component responsible for detection, tracking, and recognition of faces in video, and the audio analysis component, which operates by speaker identification. Both components have their advantages under different circumstances and we studied how to exploit the interaction between them for improved performance. Two fusion strategies are compared in our research. In the first strategy, the audio-verify-visual fusion strategy, speaker identification is used to verify the face recognition result. The second strategy, the visual-aid-audio fusion strategy, consists of using face recognition and tracking to supplement speaker identification results. By comparing the output from our system with our ground truth database, we evaluate the performance of each individual analysis component and their fusion. The results show that while the audio-verify-visual fusion strategy has slightly lower recall than the original face recognition system, it achieves the best identification precision among different algorithms. This suggests that such a strategy is suitable for applications where precision is much more critical than recall (e.g., security systems). The visual-aid-audio fusion strategy, on the other hand, generates the best overall identification performance. It outperforms either of the individual analysis components greatly in both precision and recall. This strategy is suitable to more general applications, such as, in our case, person identification in TV programs.
