Correlation filters are a standard way to solve many problems in signal processing, image processing, and computer vision. NeoFilter Labs introduces two new filter training techniques, called Average of Synthetic Exact Filters (ASEF) and Minimum Output Sum of Squared Error (MOSSE), which have produced filters that perform well on a many object detection and tracking problems. 

Typically, correlation filters are created by cropping templates out of training images; however, these templates fail to adequately discriminate between targets and background in difficult detection scenarios. More advanced methods such as Synthetic Discriminant Functions (SDF), Minimum Average Correlation Energy (MACE), Unconstrained Minimum Average Correlation Energy (UMACE), and Optimal Tradeoff Filters (OTF) improve performance by controlling the response of the correlation peak, but they only loosely control the effect of the filters on the rest of the image. ASEF and MOSSE are fundamentally new approaches to correlation filter training, which considers the entire image to image mapping known as cross-correlation. ASEF and MOSSE find filters that optimally map the input training images to user specified outputs. The goal is to produce strong correlation peaks for targets while suppressing the responses to other background objects. 

In tests, ASEF and MOSSE filters have significantly outperformed other correlation filter training techniques in object detection scenarios because of their ability to suppress background responses and therefore boosts the correct detection rate.  For example, the filters accurately detected eyes in facial imagery 96% which was 20% better that the best scores for filters based MACE and UMACE.  Additionally the unique training techniques employed by NeoFilter Labs were able distinguish between the right and left eyes which are visually similar.  Another example is person detection in video, where MOSSE resulting in 87% correct detection rate compared to UMACE which resulted in 67%.  Not only did ASEF and MOSSE outperform other correlation based approaches, but the accuracy was shown to meet or exceed well established and much slower detection algorithms including the Viola and Jones cascade detector.

NeoFilter Labs has also developed a visual tracking algorithm that is robust to appearance changes, scale changes, rotation changes, lighting changes, camera motion, and it can even detect and handle occlusions.  The algorithm is much faster than other recently published algorithms (669 fps vs 25fps on an Intel Core 2 Duo 2.4Ghz) and it can follow targets much better than any other tracker we have evaluated.  This significantly reduces the CPU load needed to track targets and therefore allows the tracker to follow multiple objects at the same time or frees system resources for other computations.  We are also looking at porting this tracker to small embedded CPUs such as ARM based architectures.


Technical Contact: 
Sales & Licensing Contact: 


Home             Technical               Publications                Videos             Products