Object detection and segmentation based on shape learning and Dynamic Bayesian network
Qinkun Xiao, Xiangjun Liu, Song Gao, Haiyun Wang · 2010
An novel object detection and segmentation algorithm is proposed based on contour information and inference. The scheme proceeds in mo stages, in the first stage, some prototypes are learned automatically from their 3D models and training images. Firstly, aiming to viewpoint invariance, a two-layer structural of multi-views is built, secondly, the corner-polygon (CP), which is combination of 3 features, is proposed for describing templates. In the second stage, the generated framework is used for detecting category-given object. A new Dynamic Bayesian network (DEN) model is built to infer detection probability, based on corresponding patch matching and inference computing, the candidate window s should be found out, then thin plate spline (TPS) technique is used to segment object from background. Experimental results demonstrate that our proposed approach is able to identify and accurately detect and segment the objects with better performance than the existing methods.