One of important conferences and gatherings in Computer Vision took place in October 2016 in Amsterdam, The Netherlands. It was the The 14th European Conference on Computer Vision – Amsterdam, The Netherlands and featured numerous interesting presentations and research developments.
Today I share with The Information Age readers and followers what I thought to be one of the best such presentations. It was about a research paper from the Swedish Linköping University’s Department of Electrical Engineering and it concerned a new method to improve visual object tracking and feature point tracking using deep convolutional feature maps with a training continuous method. This method is also related to the now widely known Generative Adversarial Networks framework, as it is based on a discriminative filter – it is actually called Discriminative Correlation Filters (DCF).
Below I provide the link to the video presentation that I found in videolectures.net as well as a link with the PDF of the research paper. As Martin Danelljan point out in the video, the paper detailed the main points described in the lecture and if there is anyone interested in knowing more about the further details of the research is free to contact the authors or send comments and suggestions.
The mathematics and signal processing in this paper is advanced a quite difficult to a wider readership. With that said, what I would like to stress is that science and engineering is full of examples where in spite of the complexity and difficulty with the theoretical frameworks behind the research there is an intuitive and easy to see application, and the common view just see its effects. I just happen to be thrilled by what kind of magic this research efforts end up achieving in the demanding but fascinating subject of Computer Vision and Artificial Intelligence.
The Videolectures.net link:
The link to the research paper with the abstract below:
Discriminative Correlation Filters (DCF) have demonstrated excellent performance for visual object tracking. The key to their success is the ability to efficiently exploit available negative data by including all shifted versions of a training sample. However, the underlying DCF formulation is restricted to single-resolution feature maps, significantly limiting its potential. In this paper, we go beyond the conventional DCF framework and introduce a novel formulation for training continuous convolution filters. We employ an implicit interpolation model to pose the learning problem in the continuous spatial domain. Our proposed formulation enables efficient integration of multi-resolution deep feature maps, leading to superior results on three object tracking benchmarks: OTB-2015 (+5.1% in mean OP), Temple-Color (+4.6% in mean OP), and VOT2015 (20% relative reduction in failure rate). Additionally, our approach is capable of sub-pixel localization, crucial for the task of accurate feature point tracking. We also demonstrate the effectiveness of our learning formulation in extensive feature point tracking experiments. Code and supplementary material are available at:
I confess to not having read fully and properly the whole paper. But I have watched with deeply interested attention the video presentation and it is an appetizer to delve into the details of the paper, even if by superficial inspection its content seems a bit intimidating. But grown up I am and it is regretful if I don’t do it.