Image Segmentation Tutorial
Image segmentation has come a long way. Using just a few simple grouping cues, one can now produce rather impressive segmentation on a large set of images. Behind this development, a major converging point is the use of graph based technique. Graph cut provides a clean, flexible formulation for image segmentation. It provides a convenient language to encode simple local segmentation cues, and a set of powerful computational mechanisms to extract global segmentation from these simple local(pairwise) pixel similarity. Computationally graph cuts can be very efficient.
In this tutorial, we will summarize current progress on graph based segmentation in four topics:
1) general graph cut framework for image segmentation: Normalized Cuts, Typical Cuts, and Min Cuts;
2) data human image segmentation, and segmentation benchmark;
3) image statistics and grouping cues: intensity, texture;
4) multi-scale graph cut.
Panel Disussion: Segmentation, Recognition, and Scene ReasoningSpeakers: Jitendra Malik, David Forsyth, Yann LeCun, Ronen Basri, Antonio Torralba, Pedro Felzenszwalb
Image segmentation and recognition are two intertwined topics. Does segmentation leads to recognition, or recognition leads to segmentation? Regardless one's philosophical stand on this question, it is undeniable a tight connection exists between them. Several proposals have emerged recently, some uses top-down recognition process to guide image segmentation, while others use bottom-up segmentation to guide object recognition. The results have been surprisingly good in their limited domain. In this panel discussion, we aim to explore a full spectrum of these possibilities.
Demonstration and Benchmark Test:
We plan to organize a demonstration session in the end of this tutorial, showing different type of segmentation algorithms, and comparing them on a common benchmark test. For the benchmark test, we plan to start with the Berkeley Segmentation Dataset and Benchmark based on human-segmented natural images.