The Study of RNN Enhanced Convolutional Neural Network for Fast Object Detection Based on the Spatial Context Multi-Fusion Features
Ning Wang, Aihua Cai, Shunsheng Zhang · 2018
In object detection, the spatial context information of the actual scene can improve the accuracy of detection while current object detection methods based on deep learning usually relies on CNNs to combine the local geometric and texture features without using these context information. In this paper, we proposal a method inspired by RNN that can extract the spatial context features and design a lighter, faster and more accurate network model for object detection that can fuse these features into multi-scale feature maps from CNNs to generate more comprehensive multi-information fusion features as inputs of localization and classification tasks subnetwork. The spatial context features extractor can be trained by joining in the BP process of the entire model, and we give the detailed optimized derivation procedures. Then we also demonstrate the validity of our model by conducting the contrastive experiments on KITTI dataset.