|
Object detection methods can be divided into two categories that are the two-stage methods with higher accuracy but lower speed and the one-stage methods with lower accuracy but higher speed. In order to inherit the advantages of both approaches, a novel dense object detector, called Path Augmented RetinaNet (PA-RetinaNet), is proposed in this paper. It not only achieves a better accuracy than the two-stage methods, but also maintains the efficiency of the one-stage methods. Specifically, we introduce a bottom-up path augmentation module to enhance the feature exaction hierarchy, which shortens the information path between lower feature layers and topmost layers. Furthermore, we address the class imbalance problem by introducing a Class-Imbalance loss, where the loss of each training sample is weighted by a function of its predicted probability, so that the trained model focuses more on hard examples. To evaluate the effectiveness of our PA-RetinaNet, we conducted a number of experiments on the MS COCO dataset. The results show that our method is 4.3 \% higher than the existing two-stage method, while the speed is similar to the state-of-the-art one-stage methods. |
|
Keywords:Object Detection; Convolutional Neural Network; Class Imbalance |
|