ECCV 2016 Daily - Wednesday

Salehe Erfanian Ebadi 16 ECCV Daily : Wednesday Presentations Foreground Segmentation via Dynamic Tree-Structured Sparse RPCA Salehe Erfanian Ebadi is a PhD student at Queen Mary University of London, working under the supervision of Professor Ebroul Izquierdo. Together they developed a video recording project with BBC called THIRA and a security project for Europe and the UK called LASIE. This project seeks to get foreground segmentation and background modeling in general video sequences. In a video, foreground can include a lot of things. For example, if a video contains a moving object such as a person walking across the scene, the person is foreground. Then, if the person puts down a piece of luggage and leaves the scene without the luggage, that’s also foreground. The same for cars moving or entering/parking/leaving the scene. The main goal is to find all of the regions that are interesting for further processing, which can include tracking, segmentation, or recognition. It can recognize the person or detect if the object left on screen contains a bomb. That’s the practical application of the project. However, sensors are very local. In addition, CCTV cameras provide poor images and color definition with high noise. The video may also have low lighting situations making it difficult for computers to detect moving objects. This can prove difficult even for humans, making it even more challenging for a computer to detect all of the parts correctly. For this work, they use heavily modified and approximated robots principal component analysis; and dynamic tree structured sparsity to get very good, crisp foreground segmentation accuracy. The method was tested on 4 different data sets, which are benchmark background subtraction data sets providing state of the art results for all four. The next step is making it real-time to make this into a product. The most challenging thing for Salehe was to actually get this published, even without using deep learning (the hype of these days, in her words). Instead, she makes an effort and prove that this model works better than deep learning. Still, she successfully published her work, e.g. with ECCV (find it in the proceedings) and also in in a couple of ICIP papers as well.