What We Talk About When We Talk About Object

In a CVPR 2010 paper, Vittorio Ferrari raised the question of "What is an object?", that is, how could we give a measure of objectness generic over classes? I think it is an interesting question which not only relates to generic object detection, but also shed light on saliency or attention analysis and segmentation tasks.

First, we give a brief introduction of Ferrari's generic object detection method. Ferrari argues that any object has at least one of three distinctive characteristics:

• a well-defined closed boundary in space;
• a different appearance from their surroundings;
• sometimes it is unique within the image and stands out as salient.2. Improvements under weakly supervised manner:

According to these three disciplines, Ferrari utilizes four image cues in the objectness measure: Multi-scale Saliency (MS), Color Contrast (CC), Edge Density (ED) and Superpixels Straddling (SS). All the parameters in those image cues are learned from a Bayesian framework. Below are some generic object detection results using Ferrari's method:

Ferrari's another paper named "Figure-ground segmentation by transferring window masks" further the discussion of generic object detection by introducing the transfer learning method.

The intuition behind the weakly supervised method is that similar windows often have similar segmentation masks. Therefore we could rely on nearest neighbors in the appearance space to transfer segmentation masks. And the appearance model consists of two gaussian mixture models (GMM), one for the foreground, one for the background. Below are some segmentation results as well as their masks: