Multimodal Mixed Conditional Random Field Model for Category-Independent Object Detection

Category-independent object detection is extremely useful for many robot vision tasks. Most existing methods rank a lot of regions by measuring their object-likeness. However, to obtain a sufficient object covering rate too many regions need to be sampled. In this paper, we present a novel method that directly detects and localizes category-independent objects. We develop a novel model which is named as “mixed robust higher-order conditional random field” model which combines 2D and 3D data into a uniform framework. A set of novel features is developed based on 2D and 3D saliency and oversegments. The potentials used in this model are computed from these features. Extensive experiments are carried out on a public RGB-D dataset. By comparison with state-ofthe- art ranking methods, the experimental results show the comparable performance of category-independent object detection without sampling a large number of extra regions.