Professional Documents
Culture Documents
Slides Xi Spia
Slides Xi Spia
Introduction
XI Peruvian Symposium on Artificial Introduction
Intelligence (SPIA)
Motivation
2
Motivation Introduction Proposed Method Experiments Conclusion
Motivation
3
Motivation Introduction Proposed Method Experiments Conclusion
Motivation
4
Motivation Introduction Proposed Method Experiments Conclusion
Motivation
a) Work
5
Motivation Introduction Proposed Method Experiments Conclusion
Introduction
• Authors usually do not consider the face behavior and head motion of
videos.
begin end
6
Motivation Introduction Proposed Method Experiments Conclusion
Introduction
Proposed Method
• This study considers the importance of face behavior and head motion to
recognize facial expressions.
• We propose to generate texture maps [1] to encode the face variations and
motion.
• We consider the Landmark [2] points to generate our texture maps.
Fig. Facial expression sequence begins with a neutral expression and proceeds to
target expression. For each frame, we compute its respective landmark points . 8
Motivation Introduction Proposed Method Experiments Conclusion
Proposed Method
Proposed Method
Grouping Points
10
Motivation Introduction Proposed Method Experiments Conclusion
Proposed Method
11
Motivation Introduction Proposed Method Experiments Conclusion
Proposed Method
12
Motivation Introduction Proposed Method Experiments Conclusion
Proposed Method
Classification
• We use transfer-learning.
• We use the pre-trained ImageNet-vgg-f model.
• The images are resized to 256x256x3.
13
Motivation Introduction Proposed Method Experiments Conclusion
Experiments
Fig. The CK+ dataset consists of 593 images sequences, 123 subjects
performed 8 facial expressions categories.
14
Motivation Introduction Proposed Method Experiments Conclusion
Experiments
15
Motivation Introduction Proposed Method Experiments Conclusion
Conclusion
• It is necessary to combine facial changes with face motion and local information to
produce a robust method.
• In this work, we prove that using texture maps is a feasible way to encode the
temporal information.
• As future work, we will extend our CNN models to improve the achieving results
and test our method on a continuous dataset.
16
Motivation Introduction Proposed Method Experiments Conclusion
References
[1] Zewei Ding, Pichao Wang, Philip O Ogunbona, and Wanqing Li. Investigation of different
skeleton89features for cnn-based 3d action recognition. In2017 IEEE International Conference on
Multimedia90& Expo Workshops (ICMEW), pages 617–622. IEEE, 2017b.
[2] Yuval Nirkin, Iacopo Masi, Anh Tran Tuan, Tal Hassner, and Gerard Medioni. On face
segmentation,92face swapping, and face perception. In2018 13th IEEE International Conference on
Automatic93Face & Gesture Recognition (FG 2018), pages 98–105. IEEE, 2018.
[3] Patrick Lucey, Jeffrey F Cohn, Takeo Kanade, Jason Saragih, Zara Ambadar, and Iain Matthews.
The69extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified
ex-70pression. In2010 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition-71Workshops, pages 94–101. IEEE, 2010.
[4] Hui Ding, Shaohua Kevin Zhou, and Rama Chellappa. Facenet2expnet: Regularizing a deep
face83recognition net for expression recognition. In2017 12th IEEE International Conference
on84Automatic Face & Gesture Recognition (FG 2017), pages 118–126. IEEE, 2017a.
[5] Nianyin Zeng, Hong Zhang, Baoye Song, Weibo Liu, Yurong Li, and Abdullah M Dobaie.
Facial86expression recognition via learning deep sparse autoencoders. Neurocomputing, 273:643–
649,872018.
[6] Mengyi Liu, Shaoxin Li, Shiguang Shan, and Xilin Chen. Au-aware deep networks for
facial97expression recognition. In2013 10th IEEE International Conference and Workshops on
Automatic98Face and Gesture Recognition (FG), pages 1–6. IEEE, 2013. 17
.