Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
4947326 | Neurocomputing | 2017 | 24 Pages |
Abstract
Research on Region Convolution Neural Network (RCNN) based object localization has recently witnessed rapid progress, but constrained by the size of the output convolution map, this method is unable to obtain exact object positions. In this paper, we present a multi-task learning approach on convolution neural network for object localization. Our model consists of 3 modules, respectively extracting shared features, generating low-level features, and fusing different levels information. We developed an algorithm for the nontrivial end-to-end training of this causal, cascaded structure. Finally, we demonstrated performance of the algorithm on the PASCAL VOC 2007 dataset and traffic scene dataset. Experiments show that our algorithm effectively and efficiently improved performance.
Related Topics
Physical Sciences and Engineering
Computer Science
Artificial Intelligence
Authors
Yan Tian, Huiyan Wang, Xun Wang,