Efficient Region Proposals Generation for Object Detection

Object detection models based on region proposals, whether those using classic methods or deep neural networks, have provided cutting-edge results and achieved an interesting speed/accuracy ratio. The state-of-the-art methods using local feature detectors to generate region proposals incorporate descriptor information or combine the detected interest points with other methods to generate proposals. In this work, we aim to explore the efficiency of proposals generated only according to the position of the generated SURF keypoints. We propose an object detection approach namely SURF-Fast-RCNN, based on FAST-RCNN framework, that integrates SURF local detector. Instead of using the proposals produced by the selective search method, each 2-combination of interest points is considered as a candidate box which we call keypoint anchors. Based on the intersection over union (IoU) value of each keypoint anchor with ground truth boxes, we discard the excess anchors and assign labels to the retained ones denoting their class labels. The retained anchors are fed into the detector for training, followed by NMS operation to suppress the overlapping boxes. Our SURF-Fast-RCNN model is trained and tested on different region proposals on the MS-COCO dataset examples. With 1000 region proposals, the performances of our model reached an accuracy of about 82% and a mean average precision of 84.04%, which are encouraging results.