DSN: A New Deformable Subnetwork for Object Detection

Although deep convolutional neural networks have achieved great success in object detection, they depend heavily on considerable training data and do not have a specific mechanism to handle related challenging problems, such as object deformation. In this paper, we design a deformable subnetwork (DSN) to introduce the deformable part-based model (DPM) in the deep object detection framework. It is effective for handling object deformation and is composed of two significant parts: the deformation coefficient part and the deformation pooling part. The deformation coefficient part is responsible for generating the deformation coefficients for each position of the input. The deformation pooling part calculates the final score for each position which takes into account its displacement penalty relative to the root position. DSN is convenient for being embedded into the most prevalent object detection frameworks such as faster-RCNN or RFCN. More importantly, it does not impair the integrity of the original framework and only causes little time consumption. We show the effectiveness of DSN via experiments on the PASCAL VOC and COCO datasets, achieving state-of-the-art results, 82.7% for PASCAL VOC and 32.1% for COCO.