Face retrieval in news video has been identified as a challenging task due to the huge variations in the visual appearance of the human face. Although several approaches have been proposed to deal with this problem, their extremely high computational cost limits their scalability to large-scale video datasets that may contain millions of faces of hundreds of characters. In this paper, we introduce approaches for face retrieval that are scalable to such datasets while maintaining competitive performances with state-of-the-art approaches. To utilize the variability of face appearances in video, we use a set of face images called face-track to represent the appearance of a character in a video shot. Our first proposal is an approach for extracting face-tracks. We use a point tracker to explore the connections between detected faces belonging to the same character and then group them into one face-track. We present techniques to make the approach robust against common problems caused by flash lights, partial occlusions, and scattered appearances of characters in news videos. In the second proposal, we introduce an efficient approach to match face-tracks for retrieval. Instead of using all the faces in the face-tracks to compute their similarity, our approach obtains a representative face for each face-track. The representative face is computed from faces that are sampled from the original face-track. As a result, we significantly reduce the computational cost of face-track matching while taking into account the variability of faces in face-tracks to achieve high matching accuracy. Experiments are conducted on two face-track datasets extracted from real-world news videos, of such scales that have never been considered in the literature. One dataset contains 1,497 face-tracks of 41 characters extracted from 370 hours of TRECVID videos. The other dataset provides 5,567 face-tracks of 111 characters observed from a television news program (NHK News 7) over 11 years. We make both datasets publically accessible by the research community. The experimental results show that our proposed approaches achieved a remarkable balance between accuracy and efficiency.

Depth-based action recognition has been attracting the attention of researchers because of the advantages of depth cameras over standard RGB cameras. One of these advantages is that depth data can provide richer information from multiple projections. In particular, multiple projections can be used to extract discriminative motion patterns that would not be discernible from one fixed projection. However, high computational costs have meant that recent studies have exploited only a small number of projections, such as front, side, and top. Thus, a large number of projections, which may be useful for discriminating actions, are discarded. In this paper, we propose an efficient method to exploit pools of multiple projections for recognizing actions in depth videos. First, we project 3D data onto multiple 2D-planes from different viewpoints sampled on a geodesic dome to obtain a large number of projections. Then, we train and test action classifiers independently for each projection. To reduce the computational cost, we propose a greedy method to select a small yet robust combination of projections. The idea is that best complementary projections will be considered first when searching for optimal combination. We conducted extensive experiments to verify the effectiveness of our method on three challenging benchmarks: MSR Action 3D, MSR Gesture 3D, and 3D Action Pairs. The experimental results show that our method outperforms other state-of-the-art methods while using a small number of projections.