Although the recent success of convolutional neural network (CNN) advances state-of-the-art saliency prediction in static images, few work has addressed the problem of predicting attention in videos. On the other hand, we find that the attention of different subjects consistently focuses on a single face in each frame of videos involving multiple faces. Therefore, we propose in this paper a novel deep learning (DL) based method to predict salient face in multiple-face videos, which is capable of learning features and transition of salient faces across video frames. In particular, we first learn a CNN for each frame to locate salient face. Taking CNN features as input, we develop a multiple-stream long short-term memory (M-LSTM) network to predict the temporal transition of salient faces in video sequences. To evaluate our DL-based method, we build a new eye-tracking database of multiple-face videos. The experimental results show that our method outperforms the prior state-of-the-art methods in predicting visual attention on faces in multipleface videos
Multiple-Face Videos with Eye Tracking fixations (MUFVET). All videos in MUFVET are with either indoor or outdoor scenes, selected from Youtube and Youku, and they are all encoded by H.264 with duration varying from 10-20 seconds. Besides, MUFVET includes two datasets – MUFVET-I and MUFVET-II. These two datasets are comprised by two non-overlapping groups of videos, each of which is viewed by totally different subjects.
We think both training and test utilize the fixations of same subjects are not rationale in existing saliency prediction works, despite videos being different. So MUFVET-I is seen as the benchmark for test, while MUFVET-II is used for training.
- Tensorflow
- Keras
- Matlab
- Python 2.7
Our | GT | Xu et al. | Jiang et al. | GBVS | Rudoy et al. | PQFT | Surprise | OBDL | |
---|---|---|---|---|---|---|---|---|---|
NSS | 4.12 | 4.21 | 3.14 | 0.97 | 1.23 | 1.42 | 0.88 | 0.88 | 1.62 |
CC | 0.74 | 0.77 | 0.61 | 0.29 | 0.33 | 0.36 | 0.22 | 0.21 | 0.30 |