Two extreme views on object attributes in the community

I have a paper on attribute-based transfer learning for object categorization in ECCV this year. So I am very curious about the views or attitudes of the community on this topic. During this ECCV, there is a one-day workshop on this topic. At the end of this workshop, there is a panel discussion about this topic among five leading researchers in the computer vision community. It turns out that there are two views on this topic which occupy the two extremes of the spectrum. The followings are summaries of their personal views on object attributes: 

• Malik doesn’t favorite attributes. He said “vision should not be hijacked by language”

• Mata doesn’t favorite attributes. He said “my dog can recognize as good as the state-of-the-art computer vision algorithms or even better without language”

• Hoiem considers attributes a way to go beyond recognition for image understanding, i.e., describing objects and scene

• Fei-Fei considers attributes as a knowledge

• Lampert considers attributes as a way to transfer knowledge to the vision system

Overall, there is neither clear definition on attributes nor consensus in the community. It is still a controversial topic. But  it may be a hot research topic in the next a few years. In this ECCV, there are three papers about attributes.

• Automatic Attribute Discovery and Characterization from Noisy Web Data 

• Idea:  mining text and image data sampled from the Internet
• Motivation: product images online are often accompanied texts describing their attributes, such as color, parts, functions, etc.

• A Discriminative Latent Model of Object Classes and Attributes
• Attribute-based Transfer Learning for Object Categorization with Zero or One Training Example (my paper)

papers: visual attribute and object class recognition

Several recent papers discussed the methods to extract visual attributes from images and/or use these attributes for object class recognition. We can view visual attributes as another type of annotation. While image annotation is applied to individual image, visual attributes are specified to an object class; image annotations are usually words, i.e., discrete values, visual attributes can be either discrete (e.g. color={red, blue, ...}) or continuous values (e.g., average size). Here are a few papers I am reading:

• Describing Objects by their Attributes, CVPR 2009
  
• Joint learning of visual attributes, object classes and visual saliency, ICCV 2009
• Learning To Detect Unseen Object Classes by Between-Class Attribute Transfer, CVPR 2009
• Learning Visual Attributes, NIPS 2007

Here is a blog about visual attribute from Tombone.