Presentation is loading. Please wait.

Presentation is loading. Please wait.

Computer Vision as an Engineering Problem. A Hierarchical Layer Model. A presentation by Amit Benbassat.

Similar presentations


Presentation on theme: "Computer Vision as an Engineering Problem. A Hierarchical Layer Model. A presentation by Amit Benbassat."— Presentation transcript:

1 Computer Vision as an Engineering Problem. A Hierarchical Layer Model. A presentation by Amit Benbassat

2 The current gridlock in computer vision research Good results in low level vision. Problems with mid-level vision and upwards. Why?

3 It is my contention that bad concepts and design are major culprits. 1. Messy thinking: Computer vision a royal mess of concepts and ideas. Too many psychological and philosophical concepts have caught hold into what is basically an engineering problem; weighing on it and complicating it unnecessarily. 2. Overshooting: The task of vision has become the task of vision and comprehension. Although an individual can see only as well as its mind can, understanding is a still a mental faculty and need not be intertwined with the more technical aspects of vision. 3. Lack of a good hierarchic modularity and design in the process: A process trying to infer the 3D attributes of an object shouldn’t have to deal with pixels.

4 Modeling computer vision as it’s defined now. Automatic inference of properties of the world from images. Failing #3 Failing #2

5 Redefining remodeling and computer vision Real time translation of real world visual input into appropriate data for a computational mental inference process.

6 Inspiration for CVM design: Network layer models. In every host there’s code for all layers. Every layer can communicate with the layers directly above and below it on the same host via interface. Every layer has a language, called a protocol, which it shares with its counterparts on other hosts. The result of this design: Corresponding layers on different hosts can converse as if the lower levels don’t exist.

7 The Network Layer model. The layers and their functions. The physical layer: Dealing with the physical medium. The data link layer: Transferring data to neighbouring stations on the networks. Flow control. The network layer: Navigating the network to discover the destination host. The transport layer: Direct communication with remote host. Flow and congestion control. The Application layer: Managing applications (Web surfing, Email etc.)

8 The flow of information layer model. Every layer is an array of stations (In theory each station is a different physical entity). There’s a one way flow of information. Higher layers contain less stations which are fed information from larger receptive fields. The arrows signify the direction of the flow of information

9 The flow of information layer model. The layers and their functions. The physical layer: Dealing with the physical medium. Translating light into digital information. The local patterns layer: Discovering local patterns. Edgel detection. General movement detection. No 3D patterns. The large patterns layer: Many types of stations. Finds larger patterns using local patterns information. Finds some ambiguous 3D patterns. Further research required to discover methods of working with non-pixel information. The mental interface layer: One or very few stations in this layer. Cleans up and patches together information from the large patterns layer before sending it to the mental module. Again, further research is required. Language of mental module must be known. Important: This is Nothing more than a blueprint for a solution. This is by no way a working model, and any design features are subject to change.

10 Conclusion. There is a problem with computer vision. A case can be made that this problem is due to fundamental faults in the design and definition of the problem. After redefinition and remodeling of the problem, a layer based solution on is suggested. The solution -though it is far from complete and may be subject to many changes- opens a window to the possibility of better computer vision based on hierarchic design.


Download ppt "Computer Vision as an Engineering Problem. A Hierarchical Layer Model. A presentation by Amit Benbassat."

Similar presentations


Ads by Google