A Dual-branch CNN Structure for Deformable Object Detection

Abstract

Object detectors based on CNN are now able to achieve satisfactory accuracy, but their ability to deal with some targets with geometric deformation or occlusion is often poor. This is largely due to the fixed geometric structure of the convolution kernel and the single inflexible network structure. In our work, we use dual branch parallel processing to extract the different features of the target area to coordinate the prediction. To further enhance the performance of the network, this study rebuilds the feature extraction module. Finally, our detector learns to adapt to a variety of different shapes and sizes. The proposed method achieves up to 81.76% mAP on the Pascal VOC2007 dataset and 79.6% mAP on the Pascal VOC2012 dataset.