 |
Center for Nonlinear Studies |
Understanding boundaries between image regions provides a wealth of information about the structure and contents of a visual scene. While the problem of image segmentation has been studied for many years, progress has often been hampered by lack of objective evaluation criteria. I will describe our recent efforts built around a large dataset of natural images which have been hand segmented by multiple human subjects. This data provides ground-truth that can be used both for benchmarking algorithm performances and as training data for machine learning based approaches. I will show that, relative to this benchmark, our ability to automatically segment natural images based on low-level cues has shown significant improvement over the last 40 years and is now drawing close to human level performance. Segmentation and boundary detection provide a firm foundation for higher-level visual tasks such as object recognition.
Host: Alexei N. Skurikhin, ISR-2