Duration: 59:55

19 views

1

We consider the problem of finding instances of visual object categories (such as a cup or a pen) in cluttered, real-world environments.  We propose a hierarchical approach, whereby we first categorize the scene (outdoors or indoors? kitchen or office?), then we use global image statistics (the gist of the image) to predict where to look within the image, and finally we run an object detector (based on boosted random fields) to localize the object, exploiting spatial constraints with other, easier-to-detect objects. We argue that this top-down approach is not only faster than standard brute