Abstract

In this paper, we investigate two new strategies to detectobjects accurately and efficiently using deep convolutionalneural network: 1) scale-dependent pooling and 2) layer-wise cascaded rejection classifiers. The scale-dependentpooling (SDP) improves detection accuracy by exploitingappropriate convolutional features depending on the scaleof candidate object proposals. The cascaded rejection clas-sifiers (CRC) effectively utilize convolutional features andeliminate negative object proposals in a cascaded man-ner, which greatly speeds up the detection while maintain-ing high accuracy. In combination of the two, our methodachieves significantly better accuracy compared to otherstate-of-the-arts in three challenging datasets, PASCAL ob-ject detection challenge, KITTI object detection benchmarkand newly collected Inner-city dataset, while being more ef-ficient.