A Modified PSRoI Pooling with Spatial Information

摘要

Position Sensitive RoI pooling in RFCN [5] is a RoI pooling that contains the position information. Each RoI rectangle will be devided into K × K bins by a regular grid. In this paper, we present a modified PSRoI pooling that contains spatial information. Every bin in PSRoI pooling is rescaled to 2×. With this proposed network, the spatial information around the bin will be added to predict the classifications. The weight outside the bin is simply set to 0.5 manually. We use ResNet101 backbone network to test our model on PASCAL VOC and MS COCO. We gain some improvement compared with RFCN [5] on the VOC and MS COCO Dataset.