| Summary: | <p>The world is full of objects: cups, phones, computers, books, and</p><p>countless other things. For many tasks, robots need to understand that</p><p>this object is a stapler, that object is a textbook, and this other</p><p>object is a gallon of milk. The classic approach to this problem is</p><p>object recognition, which classifies each observation into one of</p><p>several previously-defined classes. While modern object recognition</p><p>algorithms perform well, they require extensive supervised training:</p><p>in a standard benchmark, the training data average more than four</p><p>hundred images of each object class.</p><p>The cost of manually labeling the training data prohibits these</p><p>techniques from scaling to general environments. Homes and workplaces</p><p>can contain hundreds of unique objects, and the objects in one</p><p>environment may not appear in another.</p><p>We propose a different approach: object discovery. Rather than rely on</p><p>manual labeling, we describe unsupervised algorithms that leverage the</p><p>unique capabilities of a mobile robot to discover the objects (and</p><p>classes of objects) in an environment. Because our algorithms are</p><p>unsupervised, they scale gracefully to large, general environments</p><p>over long periods of time. To validate our results, we collected 67</p><p>robotic runs through a large office environment. This dataset, which</p><p>we have made available to the community, is the largest of its kind.</p><p>At each step, we treat the problem as one of robotics, not disembodied</p><p>computer vision. The scale and quality of our results demonstrate the</p><p>merit of this perspective, and prove the practicality of long-term</p><p>large-scale object discovery.</p> === Dissertation
|
|---|
