Object detection and recognition: from saliency prediction to one-shot trained detectors
Adrià Recasens Continente · 2014
[ANGLES] Computer vision capabilities have started to become available in smart devices this last years. The rapid growth of the smartphone world along with the big advance of the computer vision field in the last years make possible nowadays to bring computer vision to everyday's mobile devices. DetectMe is one of the firsts systems to bring object detectors to everyone's mobile device. This paradigm shift generates new challenges and questions: this project wants to answer some of this questions as well as give some future lines of work to overcome this challenges. On one hand, the aim of this project is to answer a short question: can we train good detectors with only one example? Section 3 will analyze this issue as well as point out aside questions that appear when the main question is trying to being answered. A positive answer to this question as well as some hints on what makes an object a good example would improve the user experience for those who are using computer vision systems in mobile devices. On the other hand, we will also attack a classical problem in computer vision: where people look when they are looking at a picture? The recent development of the Convolutional Neural Networks and its outstanding capabilities to explain visual information help to improve the performance on the saliency models. In section 4, a new saliency model is presented and discussed. Results show that our saliency model outperforms the state-of-the-art saliency models on the MIT 1001 dataset. Some future research lines are also drawn to improve the model as well as generate more saliency data to work with. To sum up, this project doesn't want to be a closed project. It wants to answer some questions while pointing out potential future lines of research to find a more complete answer.