Affine Stable Characteristic based sample expansion for object detection

Generating better object model from automatic expanded samples is an effective approach to improve the performance of object detection. However, most existing methods either don't work well with limited relevance images in corpus, or result in redundant features and the decrease of detection speed. In this paper, we propose a novel method called Affine Stable Characteristic to generate an object feature model using only one object sample. By integrating affine simulation with stable characteristic mining, a compact and informative object model is generated with high robustness to viewpoint and scale transformations. For characteristic mining, two new notions, Global Stability and Local Stability, are introduced to calculate the robustness of each object feature from complementary hierarchies. And they are combined to generate the final object feature model. Experiments show that our novel method is capable of detecting objects in various geometric and photometric transformations, while only acquiring one sample image. In a compiled dataset composed of many famous test sets, the detection accuracy can be improved 35.8% compared with traditional methods at rapid on-line speed. The proposed approach can also be well generalized to other content analysis tasks.