Multiple scale sharing faster-RCNN

Abstract

Small object detection is a challenging task in the field of computer vision because the objects are always of low resolution in the original image and can be easily affected by noise. The state-of-the-art Faster RCNN object detector has good capacity of detecting large objects while small object detection is not one of its advantages. This thesis presents a novel object detector Multi-Scale Sharing Faster-RCNN (MSS-FRCNN) to solve the problem of poor detection performance of small objects by Faster RCNN. We find that upsampling the input image can benefit the small object detection performance. So MSS-FRCNN takes two images with different scales as input and then uses the two feature maps extracted from two images for RoI generation independently. Finally, the model merges the two feature map for classification and bounding box regression. We test our model with two datasets Tsinghua-Tencent 100k and Pascal VOC 07+12. The result demonstrates that MSS-FRCNN can outperform original Faster RCNN in small object detection.