Object and Contour Detection with an Architecture-Fusion Network
Keqing Jian, Shenglin Gui · 2021 IEEE 33rd International Conference on Tools with Artificial Intelligence (ICTAI) · 2021
We present a new, simple, flexible framework for object and contour detection called Architecture-Fusion Network (AFN), which fuses two basic deep learning architecture by sharing partly common stages to detect objects and their contours simultaneously and does not complicate their respective network architectures. In this paper, we train four AFN instances end-to-end on the modified PASCAL VOC 2012 dataset and effectively reduce the misrecognition of both background edges and internal edges inside objects without compromising object detection precision. Compared with each network itself, the tests on the refined PASCAL-val dataset show our instances increase Optimal Dataset Scale (ODS) by a range of 1.5% to 5.3% on object contour detection with object detection precision fluctuating slightly. Furthermore, each AFN instance has fewer parameters than the state-of-the-art network RCN. Compared with RCN on the same training dataset(RCN-VOC), the best Average Precision (AP) of our AFN instances could be up to 9.7% higher while its ODS descends only 0.9%.