Object Retrieval And Detection

We consider the problem of object retrieval, object detection and the combination of two. Given an input image, by searching or detecting the object appear in the image, the system can obtain sematic information from the image and in turn understand the input image. To solve this problem, we first develop two kinds of object retrieval system, namely BoW-based system and keypoint-matching-based system. In addition, we propose a logo detection system enabling multi-kinds of logos to be detected simultaneously. Finally, based on logo detection system, we propose a first detection then retrieval scheme to solve small object retrieval task to some extent.The contribution of our work can be summarized as follows:Firstly, we design and implement a BoW model based object retrieval system. Several parts of the system, e.g. codebook training, kernel function and indexing scheme are improved. Final experimental result on publicly available datasets shows that our system outperforms several existing methods;Secondly, we design and implement a keypoint-matching-based object retrieval system to tackle the problem of quantization error and symmetric scoring in BoW model. While keeping the memory and time efficiency at the same level compared to BoW model based systems, the experimental result on publicly available datasets shows that this system significantly improves the state-of-art object retrieval system based on BoW model.Thirdly, we propose a scalable logo detection system which enables multi-kinds of logos to be detected simultaneously. In this system, a tree-based shape descriptor is introduced to describe both the appearance and spatial information of the picture. Also, an invert indexing scheme is adopted to index multi-kinds of pre-trained logo models and makes it possible to detect multi-kinds of logos simultaneously. Final result on three publicly available datasets shows that this system out-performs several state-of-art logo detection algorithms.