Name
Abstract | ||
|---|---|---|
The performance of face localization system often degenerates when occlusions occur. In this paper, we propose Occlusion Deep Alignment Network(ODAN), which is robust to occlusion. Occlusion detection and rough face alignment are employed simultaneously, where the occlusion detection of landmarks is treated as a regression problem. Based on occlusion information, localization offset scaling is employed to transform the difference between the groundtruth and the rough landmark estimation to another linear space to generate target positions for precise face alignment, reducing the difficulty of the regression task. Finally, the outputs of the proposed model are transformed back to Euclidean Space. Experiments on two public datasets show that the proposed method achieves state-of-art performance under severe occlusion both in accuracy and efficiency. |
Year | DOI | Venue |
|---|---|---|
2019 | 10.1109/VCIP47243.2019.8966009 | 2019 IEEE Visual Communications and Image Processing (VCIP) |
Keywords | Field | DocType |
Face alignment,occlusion detection,difference transformation,regression | Computer vision,Occlusion detection,Occlusion,Regression,Computer science,Linear space,Euclidean space,Artificial intelligence,Landmark,Scaling,Offset (computer science) | Conference |
ISSN | ISBN | Citations |
1018-8770 | 978-1-7281-3724-7 | 0 |
PageRank | References | Authors |
0.34 | 12 | 5 |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Yang Li | 1 | 0 | 0.68 |
| Chang Shu | 2 | 0 | 0.34 |
| Ning Zhou | 3 | 0 | 0.34 |
| Jianyu Hong | 4 | 0 | 0.34 |
| Xiaofeng Li | 5 | 3 | 1.07 |