SDM : Squeeze and Excitation Deformable Mask-RCNN

초록

In this paper, we propose an instance-level semantic segmentation method for object detection. Squeeze and excitation deformable Mask-RCNN (SDM) improves detection by adding additional modules based on Mask-RCNN. SDM extracts informative features from the CNN architecture to obtain global information which can be acquired by squeeze and excitation. In general, the size and shape of objects are variable while the receptive field of a general CNN architecture is fixed. It is possible to detect more precisely by applying appropriate receptive field to object using deformable convolution.