Abstract

In this paper, we propose an approach that exploits ob-ject segmentation in order to improve the accuracy of objectdetection. We frame the problem as inference in a MarkovRandom Field, in which each detection hypothesis scoresobject appearance as well as contextual information usingConvolutional Neural Networks, and allows the hypothesisto choose and score a segment out of a large pool of ac-curate object segmentation proposals. This enables the de-tector to incorporate additional evidence when it is avail-able and thus results in more accurate detections. Our ex-periments show an improvement of 4.1% in mAP over theR-CNN baseline on PASCAL VOC 2010, and 3.4% over thecurrent state-of-the-art, demonstrating the power of our approach.