So the essential problem we want to solve with our tool is to automatically label an image. We want to know, In other words, want to label an image, namely, assign an object label to each pixel. This labeling task is basis for variety of applications of geospatial image data. Traditionally, this task mainly be done with manual interpretation and digitization of image. But you can imagine such manual manipulation is extremely time-consuming and label-intensive. That is the why we want to bulid tool which can automatically labeling the image.
However, this is a very challenge task, especially when spatial resolution of image data gets higher and higher these days. With the increase of spatial resolution, traditionally-used spectra feature from pixel-level become extremely unstable for the labeling task. As can you can see, the pixel of a building is essential no different from a pixel on the road. If you rely on feature from single pixel level, there’s no hope you can achieve satisfactory result.
Problems of existing technologies have motivated us to develop new solution to the labeling task. So basically, we want to develop
To achieve these objectives, we proposed a framework and a series of techiniques for efficient implementation of this framework
SmartOBIA: Next Generation of Object Based Image Analysis Yuanming Shu, Shuo Tan, Yang Gao, Jonathan Li University of Waterloo
Background • Automatic labeling (classification) of geospatial image data labeling
Problems of Existing Solutions • Not SMART enough! – Too many parameters to tune – Not accurate enough for complex scenes – Not robust for different scenes • Difficult for even experienced image analyst to use!
Our Objectives • A tool that can be easily used by non-expert users • A tool that is accurate, robust, and fast for various applications
Our Solutions • Smart Object based Image Analysis (SmartOBIA) – Label constraint image segmentation – Multi-level object classification – Object model learning with weakly-labeled and limited sample data • Patented!