Efficient Face Detection through Deep Learning

Originally presented at "The latest in AI @Bol.com", May 8th 2019 (https://www.meetup.com/bol-com-techlab/events/260444588/).

Face analysis has recently become a topic of interest for both press and science due to the implications that this technology has for our own life and privacy. Indeed, our face is inherently related to our identity. This information can be used for identification purposes, such as unlocking our phone, but it can also be used to extract personal information, such as age, gender and emotion, for commercial or statistical purposes. Therefore, given the delicate aspects of this technology, the problem has to be treated carefully. From a Machine Learning point of view, face analysis can be defined as, given an image, the extraction of necessary features to detect a face and then to identify personal information from it. Face detection needs to be the first step of any face analysis system, while auxiliary tasks such as age, gender, pose, smile classification, or even identification can follow from the extracted detection. How do we properly detect a face? How do we learn these face-related features? In this talk, I will show how Deep Learning has revolutionized computer vision for face detection and I will review the common architectures that can be used to tackle this problem. In particular, I will speak about the Single-Shot Multibox Detector (SSD) architecture as well as the implications of varying image resolutions on the detection of faces.

Transcript

1. WWW.SIGHTCORP.COM Robert-Jan Bruintjes
2. WWW.SIGHTCORP.COM Sightcorp is an AI spin-off from the University of Amsterdam specialising in face analysis technology.
3. WWW.SIGHTCORP.COM 25 countries 5 continents
4. WWW.SIGHTCORP.COM Sightcorp R&D: Deep Computer Vision State-of-the-art CV Improve products Tailormade solutions Experiments with new techs
5. WWW.SIGHTCORP.COM Sightcorp R&D: Deep Computer Vision State-of-the-art CV Improve products Tailormade solutions Experiments with new techs
6. WWW.SIGHTCORP.COM
7. WWW.SIGHTCORP.COM Efficient Face Analysis through Deep Learning
8. WWW.SIGHTCORP.COM Face Analysis
9. WWW.SIGHTCORP.COM Anonymous Audience Analytics
10. WWW.SIGHTCORP.COM
11. WWW.SIGHTCORP.COM
12. WWW.SIGHTCORP.COM People Count
13. WWW.SIGHTCORP.COM Customer Satisfaction
14. WWW.SIGHTCORP.COM Face Analysis API
15. WWW.SIGHTCORP.COM Technologies
16. WWW.SIGHTCORP.COM Technologies
17. WWW.SIGHTCORP.COM Today
18. WWW.SIGHTCORP.COM Warning:
19. WWW.SIGHTCORP.COM Ref: http:/ /mmlab.ie.cuhk.edu.hk/projects/WIDERFace/ The problem: face detection
20. WWW.SIGHTCORP.COM Traditional approaches Viola-Jones [1]: contrasting regions Ref: By Soumyanilcsc - Own work, CC BY-SA 4.0, https:/ /commons.wikimedia.org/w/index.php?curid=37353327
21. WWW.SIGHTCORP.COM Ref: http:/ /mmlab.ie.cuhk.edu.hk/projects/WIDERFace/ The problem: non-frontal faces
22. WWW.SIGHTCORP.COM Ref: www.rsipvision.com/exploring-deep-learning/ Convolutional Neural Networks (CNN) = ?% face
23. WWW.SIGHTCORP.COM Convolutional Neural Networks (CNN)
24. WWW.SIGHTCORP.COM Convolutional Neural Networks (CNN)
25. WWW.SIGHTCORP.COM Convolutional Neural Networks (CNN) Final layer = • Complex • Compact
26. WWW.SIGHTCORP.COM Convolutional Neural Networks (CNN) Final layer = • Complex • Compact Receptive field
27. WWW.SIGHTCORP.COM Face detection: simple CNN
28. WWW.SIGHTCORP.COM Face detection: simple CNN 3x3 convolution predict C + coords
29. WWW.SIGHTCORP.COM Face detection: “YOLO” CNN [2]
30. WWW.SIGHTCORP.COM Face detection: “YOLO” CNN [2] 3x3 convolution predict C + 2x coords & conf
31. WWW.SIGHTCORP.COM Multiscale •
32. WWW.SIGHTCORP.COM Multiscale
33. WWW.SIGHTCORP.COM Multiscale: “SSD” CNN [3]
34. WWW.SIGHTCORP.COM Multiscale: “SSD” CNN [3]
35. WWW.SIGHTCORP.COM Multiscale: “SSD” CNN [3]
36. WWW.SIGHTCORP.COM Multiscale: “SSD” CNN [3] 3x3 convolution predict Cx coords & conf
37. WWW.SIGHTCORP.COM Ref: SSD: Single Shot MultiBox Detector by Wei Liu et al. 2015 Multiscale: “SSD” CNN [3]
38. WWW.SIGHTCORP.COM SSD: important parameters [4] Ref: S3FD: Single Shot Scale-invariant Face Detector by Zhang et al. 2017 Overwhelming negatives
39. WWW.SIGHTCORP.COM One step further: Feature Pyramid Networks [5] Ref: Feature Pyramid Networks for Object Detection, Tsung-Yi Lin et al. 2016
40. WWW.SIGHTCORP.COM Other methods Faster R-CNN [6] / Mask R-CNN [7] • Two-stage pipeline • Slower but more accurate R-FCN [8] • Incorporate translation variance • Much more complex Ref: Ren et al. 2015 [7]
41. WWW.SIGHTCORP.COM State of the art FPN FPN R-CNN R-CNN R-FCN R-CNN + R-FCN
42. WWW.SIGHTCORP.COM Conclusion
43. WWW.SIGHTCORP.COM We’re hiring! Junior / Senior Deep Learning Engineer • Heart of AI @ Science Park Amsterdam • Work on SOTA CV every day • See your models in action • Attend international conferences sightcorp.com/jobs sightcorp.com/blog
44. WWW.SIGHTCORP.COM References 1. Viola, P., & Jones, M. (2001). Rapid object detection using a boosted cascade of simple features. CVPR (1), 1, 511-518. 2. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 779-788). 3. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C. Y., & Berg, A. C. (2016, October). SSD: Single shot multibox detector. In European conference on computer vision (pp. 21-37). Springer, Cham. 4. Zhang, S., Zhu, X., Lei, Z., Shi, H., Wang, X., & Li, S. Z. (2017). S3FD: Single shot scale-invariant face detector. In Proceedings of the IEEE International Conference on Computer Vision (pp. 192-201). 5. Lin, T. Y., Dollár, P., Girshick, R., He, K., Hariharan, B., & Belongie, S. (2017). Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 2117-2125). 6. Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems (pp. 91-99). 7. He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask R-CNN. In Proceedings of the IEEE international conference on computer vision (pp. 2961-2969). 8. Dai, J., Li, Y., He, K., & Sun, J. (2016). R-FCN: Object detection via region-based fully convolutional networks. In Advances in neural information processing systems (pp. 379-387). Slides: http://bit.ly/sightcorpfacedetection