A new fast object detection architecture combing manually-designed feature and CNN

We proposed a new fast object detection architecture based on region, which consists of two stages, using manually-designed features and convolutional neural network (CNN) respectively. The first stage is generating many initial proposal windows, and then do objectness measure and ranking, to reduce the quantity of proposal windows. In this stage, we apply four manually-designed objectness features that take much less time to compute. To do the ranking, we apply methods based on ranking SVM and non-max suppression. The second stage is that Deep CNN classifies the candidate windows into special-category and makes the final detection results. Since every window's CNN features computation time consumption is very expensive, we reduce the large set of initial proposal window in an image to small set of candidate windows to reduce the total time of computing CNN features of an image, and thus save the detection run-time. The experiment shows our proposed method achieved the comparable result on VOC 2007 with the state-of-the-art with about half time consumption.