Answers are grounded in this paper and cite their sources.
Object detection performance, as measured on the canonical PASCAL VOC, has plateaued in the last few years. The best-performing methods are ensemble systems that typically combine multiple low-level image with high-level context. In this paper, we propose a simple and detection algorithm that improves mean average precision (mAP) by more 30% relative to the previous best result on VOC 2012---achieving a mAP of53.3%. Our approach combines two key insights: (1) one can apply high-capacity neural networks (CNNs) to bottom-up region proposals in order to and segment objects and (2) when labeled training data is scarce, pre-training for an auxiliary task, followed by domain-specific-tuning, yields a significant performance boost. Since we combine region with CNNs, we call our method R-CNN: Regions with CNN features. We compare R-CNN to OverFeat, a recently proposed sliding-window detector on a similar CNN architecture. We find that R-CNN outperforms OverFeat by large margin on the 200-class ILSVRC2013 detection dataset. Source code for complete system is available at http://www.cs.berkeley.edu/~rbg/rcnn.
No takes yet. Share an insight, caveat, or question.
Girshick et al. (2013) studied this question.