Learning to focus for object proposals

Object proposal generators address the wasteful exhaustive search of the sliding window scheme in visual object detection and have been shown effective. However, the number of candidate windows is still large in order to ensure full coverage of potential objects. This paper presents a complementary technique that aims to work with any proposal generating system, amending the workflow from “propose-assess” to “propose-adjust-assess”. The adjustment serves as an auto-focus mechanism for the system and reduces the number of object proposals to be processed. The auto-focus is realized by two learning-based transformation models, one translating and the other deforming the windows towards better alignments of the objects, which are trained for identifying generic objects using image cues. Experiments on reallife image data sets show that the proposed technique can reduce the number of proposals without loss of performance.