An Improved Approach for Object Proposals Generation

Abstract

The objectness measure is a significant and effective method used for generic object detection. However, several object detection methods can achieve accurate results by using more than 1000 candidate object proposals. In addition, the weight of each proposal is weak and also cannot distinguish object proposals. These weak proposals have brought difficulties to the subsequent analysis. To significantly reduce the number of proposals, this paper presents an improved generic object detection approach, which predicts candidate object proposals from more than 10,000 proposals. All candidate proposals can be divided, rather than preclassified, into three categories: entire object, partial object, and nonobject. These partial object proposals also display fragmentary information of the objectness feature, which can be used to reconstruct the object boundary. By using partial objectness to enhance the weight of the entire object proposals, we removed a huge number of useless proposals and reduced the space occupied by the true positive object proposals. We designed a neural network with lightweight computation to cluster the most possible object proposals with rerank and box regression. Through joint training, the lightweight network can share the features with other subsequent tasks. The proposed method was validated using experiments with the PASCAL VOC2007 dataset. The results showed that the proposed approach was significantly improved compared with the existing methods and can accurately detect 92.3% of the objects by using less than 200 proposals.