Towards improving object detection

Abstract

Object detection is important in image understanding and analysis. The object detection method usually uses detection using a bounding box, and assigns a class label to each box. First, we introduce about salient region segmentation, which is the first work published at a conference, is introduced. The main idea is to represent a saliency map of an image as a linear combination of high-dimensional color space where salient regions and backgrounds can be distinctively separated. This is based on an observation that salient regions often have distinctive colors compared to the background in human perception, but human perception is often complicated and highly nonlinear. By mapping a low dimensional RGB color to a feature vector in a high-dimensional color space, we show that we can linearly separate the salient regions from the background by finding an optimal linear combination of color coefficients in the high-dimensional color space. Our high dimensional color space incorporates multiple color representations including RGB, CIELab, HSV and with gamma corrections to enrich its representative power. To further improve the performance of our saliency estimation, our second key idea is to utilize relative location and color contrast between superpixels as features and to resolve the saliency estimation from a trimap via a learning-based algorithm. The additional local features and learning-based algorithm complement the global estimation from the high-dimensional color transform-based algorithm. This work has more than 150 citations so far, and the idea is used in various computer vision fields. However, as more and more image understanding techniques using deep learning were published and achived the state-of-the-art performance, many researchers made great efforts to apply deep learning to object detection. Object detection is composed of various structures, but the most important part is the backbone network, which extracts image features. In this thesis, we introduce a simple but effective backbone network, the deep pyramidal residual network (PyramidNet), to extract a better image features compared to residual networks. The main contribution of PyramidNet is that it maximizes the regularization ability by increasing the dimension gradually for each layer. In this thesis, we describe the novel building block in residual networks, which is our main contribution during this research. We also introduce the research about region proposal networks, one of the most important parts of the object detection architecture. We successively generate proposals by using multiple RPNs, which helps detect regions that are hard to detect with a single RPN owing to the limitations of current algorithms, such as non-maximum suppression and smooth $l_1$-loss regression. Experimental results on the PASCAL VOC and MS COCO datasets showed that the detection performance improves with the proposed StackRPN when the number of RPNs is increased with a comparable inference time and memory. Finally, we conduct experiments to maximize object detection performance by synthesizing the contents of research during the Ph.D. As a result of experiments using PyramidNet, and StackRPN, the performance of COCO dataset is significantly improved compared to existing detection methods.