Abstract

Ob ject discovery algorithms group together image regions that originate from the same ob ject. This process is effective when the input collection of images contains a large number of densely sampled views of each ob ject, thereby creating strong connections between nearby views. However, existing approaches are less effective when the input data only provide sparse coverage of ob ject views. We propose an approach for ob ject discovery that addresses this prob- lem. We collect a database of about 5 million product images that capture 1.2 million ob jects from multiple views. We represent each region in the input image by a “bag” of database ob ject regions. We group input re- gions together if they share similar “bags of regions.” Our approach can correctly discover links between regions of the same ob ject even if they are captured from dramatically different viewpoints. With the help from these added links, our proposed approach can robustly discover ob ject instances even with sparse coverage of the viewpoints.