Cycle-Spinning Convolution for Object Detection

Deep convolutional networks are prominently used in object detection tasks due to their notable performances. These typically have pooling layers following the convolution, which effectively subsamples convolution output, potentially introducing aliasing. An aliased signal emerging earlier inevitably propagates throughout network and such distortion prevents getting best performance out of a network. In this study, we propose integrating cycle-spinning (CS) into layer more robust pipeline. CS makes use shift-variant characteristics pooling, where each shifted version image results different output. The combines these outputs after unshifting them remove artifacts alleviate distortion. proposed method does not introduce any additional trainable parameters can be straightforwardly integrated layers. experimental show that algorithms DOTA dataset an up 5.5% increase mAP values.