Name
Abstract | ||
|---|---|---|
A robust facial landmark detection framework is proposed in this paper, which can be trained in an end-to-end fashion and has achieved promising detection accuracy in Grand Challenge of 106-p Facial Landmark Localization. Firstly, in order to deal with challenging cases (e.g. large pose, exaggerated expression, non-uniform lighting and occlusion), a four-stage hourglass (HGs) structure is used as the backbone while a novel hierarchical block is designed to replace the standard residual block in the original HGs. Then in order to prevent the accuracy loss by the coordinates quantization, a novel function named dual soft argmax is designed for mapping the heatmap response to final coordinates. Besides, for data augmentation cutout is used and proved to be effective to the partially occluded cases. The model is trained from the beginning, and finally the best result 83.076% for AUC is achieved on the validation set. |
Year | DOI | Venue |
|---|---|---|
2019 | 10.1109/ICMEW.2019.00130 | 2019 IEEE International Conference on Multimedia & Expo Workshops (ICMEW) |
Keywords | Field | DocType |
Improved Hourglass,Dual Soft Argmax (DSA),Facial Landmark Detection | Computer vision,Residual,Hourglass,Pattern recognition,Computer science,Artificial intelligence,Quantization (signal processing),Landmark | Conference |
ISSN | ISBN | Citations |
2330-7927 | 978-1-5386-9215-8 | 0 |
PageRank | References | Authors |
0.34 | 2 | 3 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Shenqi Lai | 1 | 63 | 5.67 |
| Zhenhua Chai | 2 | 12 | 6.59 |
| Xiaoming Wei | 3 | 56 | 7.33 |