摘要
In this paper, we propose a semantic propagation network model that incorporates semantic co-occurrence, absolute position context, relative position context, and appearance into a unified framework. Multi-class object labeling is modeled as a sequential process, which is formulated with the particle filter. We modify the particle filter and extend it to static data in image. A salient object is first labeled using the appearance and absolute position context, which serves as the initialization of particle filter. The semantic information of pre-labeled objects is propagated to the subsequent objects by semantic association, which is modeled by the particle sampling. Then, the appearance and relative position context are combined to infer the next object's category. This sequential process iterates until all the objects are labeled. To represent spatial context accurately and robustly, we propose two novel spatial context descriptors. They capture the globally spatial distribution of object regions and are robust to the changes of object shapes, structures, and scales. Experimental results demonstrate the effectiveness of our method.
| 源语言 | 英语 |
|---|---|
| 页(从-至) | 1003-1018 |
| 页数 | 16 |
| 期刊 | Neural Computing and Applications |
| 卷 | 24 |
| 期 | 5 |
| DOI | |
| 出版状态 | 已出版 - 4月 2014 |
学术指纹
探究 'Semantic propagation network with robust spatial context descriptors for multi-class object labeling' 的科研主题。它们共同构成独一无二的学术指纹。引用此
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver