Object Detection Networks on Convolutional Feature Maps

摘要

Most object detectors contain two important components: a feature extractorand an object classifier. The feature extractor has rapidly evolved withsignificant research efforts leading to better deep convolutionalarchitectures. The object classifier, however, has not received much attentionand many recent systems (like SPPnet and Fast/Faster R-CNN) use simplemulti-layer perceptrons. This paper demonstrates that carefully designing deepnetworks for object classification is just as important. We experiment withregion-wise classifier networks that use shared, region-independentconvolutional features. We call them "Networks on Convolutional feature maps"(NoCs). We discover that aside from deep feature maps, a deep and convolutionalper-region classifier is of particular importance for object detection, whereaslatest superior image classification models (such as ResNets and GoogLeNets) donot directly lead to good detection accuracy without using such a per-regionclassifier. We show by experiments that despite the effective ResNets andFaster R-CNN systems, the design of NoCs is an essential element for the1st-place winning entries in ImageNet and MS COCO challenges 2015.