Multi-layer Tree Structured Model with Sparse Coded Clustering for Facial Landmark Localization
| Autor: | Yu-Ming Kuo, 郭煜民 |
|---|---|
| Rok vydání: | 2016 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 104 We propose the Multi-Layer Tree Structured Model (MTSM) with sparse coded part sharing for facial landmark localization. The MTSM consists of three component TSMs (Tree Structured Models), namely coarse TSM (c-TSM), median TSM (m-TSM) and refined TSM (r-TSM). Different from the Regressive TSM (RTSM) that uses two components, c-TSM and r-TSM, the MTSM has an additional m-TSM that can block out the false positives generated by the c-TSM at the coarse search phase. m-TSM is built on lower resolution with few parts than those of the r-TSM, and can therefore process the c-TSM detected candidates in higher speed. Because most false positives are filtered out by the m-TSM, fewer candidates are left to be processed by the relatively complex r-TSM, shortening the overall process time. We also propose the sparse coded part sharing as a new version of TSM-based landmark localization. All parts are processed by dictionary learning so that the sparse coded representations of part features can be defined. The parts are clustered using their sparse coded representations, and novel tree structures are defined using the parts across different clusters. This approach yields the Sparse Coded TSM (SC-TSM). We compare the MTSM, the SC-TSM, the RTSM and other contemporary approaches on four benchmark databases, Helen, AFW, ALFW and LFPW, and show that the performance of the MTSM and SC-TSM are competitive. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|