Deng, JiaTeng, Yun2019-09-042026-09-292019-09-042026-09-292019-05-062019-09-04http://arks.princeton.edu/ark:/88435/dsp011831cn814https://theses-dissertations.princeton.edu/handle/88435/dsp011831cn814CornerNet is a new approach to object detection that involves predicting bounding boxes as paired top-left and bottom-right keypoints. Having outperformed all existing one-stage detectors on COCO, CornerNet demonstrates that anchor boxes are not necessary, or even desirable. One major drawback of keypoint-based methods is that the improved accuracy comes at a high processing cost, and in its current state, CornerNet is prohibitively slow in applications requiring real-time detection. We address CornerNet’s inefficiency by using smaller feature maps 1/64 the size of the input image, replacing the residual module of the Hourglass backbone with a depthwise fire module, and re-implementing corner pooling to make better use of GPU parallelism. Our new lightweight CornerNet runs at 30ms on a GTX 1080Ti and achieves 34.4 AP on COCO, outperforming YOLOv3.application/pdfenPrinceton University Senior Theses