Bibliography:
Christopher F. Herot. Graphical input through machine recognition of sketches. Proceeding of the SIGGRAPH '76 Proceedings of the 3rd annual conference on Computer graphics and interactive techniques. pp. 97-102. 1976.
Summary:
This paper talks about different approaches to inferring and identifying sketches. The author tries to explain the need for finding a balance between unwieldy fully automated systems and practical ones that rely heavily on the user. Three approaches to recognizing sketches are discussed.
The HUNCH system, tries to identify sketches, just by using the syntax of drawing, without any semantic knowledge. The system is a collection of FOTRAN programs, that identify skethes, at different levels of interpretation. STRAIT, is a program that identifies corners as a function of speed. CURVIT, used the outpu of STRAIT and identified curves by fitting them to B-Splines. Both STRAIT and CURVIT worked well for some users, but not all. The focus shifted to improving STRAT's latching algorithm by comparing candidate fits in 3D instead of 2D projections. Overtracing was another problem that had to be dealt with.
The next section focusses on using context, to improve sketch recognition. The idea was, if some context information can be provided, then the syntax of sketching can be combined with this, to provide meaningful inferences of the sketch. This approach was top-down, starting from a contextual interpretation, moving down the HUNCH system interpretations and looking for matches. CONNIVER language, was one of the earliest to try this approach, but was very resource heavy and hence made it hard to construct the knowledge base.
Lookings at the merits of the two approaches above, the author suggests that the field should move towards a 'interactive system', that relies on both, automatic recognition and user interpretation. The system consists of a database and a set of programs to manipulate it. The database is dynamically updated, as new interpretations are made. Three types of programs, manipulate the database. These are 1) Inference programs that interpret the data. 2) Display programs that display the various interpretations on multiple devices. 3) Manipulation programs, that allows the user to directly manipulate the database contents. The line/curve identifier (such as the one in HUNCH) using this system, does real time recognition of sketches (with context). But the user also modifies the parameters of the programs, inorder to get more satisfactory results.
Discussion:
This paper gives a nice overview on why sketch recognition is complex without user intervention and contextual information. I think contextual information is key in sketch recognition systems, because we humans interpret similar sketches in very different ways based on contextual information.
As seen in more recent papers, seems like the research is moving towards interactive systems, that take in both user feedback and automated recognition.
Christopher F. Herot. Graphical input through machine recognition of sketches. Proceeding of the SIGGRAPH '76 Proceedings of the 3rd annual conference on Computer graphics and interactive techniques. pp. 97-102. 1976.
Summary:
This paper talks about different approaches to inferring and identifying sketches. The author tries to explain the need for finding a balance between unwieldy fully automated systems and practical ones that rely heavily on the user. Three approaches to recognizing sketches are discussed.
The HUNCH system, tries to identify sketches, just by using the syntax of drawing, without any semantic knowledge. The system is a collection of FOTRAN programs, that identify skethes, at different levels of interpretation. STRAIT, is a program that identifies corners as a function of speed. CURVIT, used the outpu of STRAIT and identified curves by fitting them to B-Splines. Both STRAIT and CURVIT worked well for some users, but not all. The focus shifted to improving STRAT's latching algorithm by comparing candidate fits in 3D instead of 2D projections. Overtracing was another problem that had to be dealt with.
The next section focusses on using context, to improve sketch recognition. The idea was, if some context information can be provided, then the syntax of sketching can be combined with this, to provide meaningful inferences of the sketch. This approach was top-down, starting from a contextual interpretation, moving down the HUNCH system interpretations and looking for matches. CONNIVER language, was one of the earliest to try this approach, but was very resource heavy and hence made it hard to construct the knowledge base.
Lookings at the merits of the two approaches above, the author suggests that the field should move towards a 'interactive system', that relies on both, automatic recognition and user interpretation. The system consists of a database and a set of programs to manipulate it. The database is dynamically updated, as new interpretations are made. Three types of programs, manipulate the database. These are 1) Inference programs that interpret the data. 2) Display programs that display the various interpretations on multiple devices. 3) Manipulation programs, that allows the user to directly manipulate the database contents. The line/curve identifier (such as the one in HUNCH) using this system, does real time recognition of sketches (with context). But the user also modifies the parameters of the programs, inorder to get more satisfactory results.
Discussion:
This paper gives a nice overview on why sketch recognition is complex without user intervention and contextual information. I think contextual information is key in sketch recognition systems, because we humans interpret similar sketches in very different ways based on contextual information.
As seen in more recent papers, seems like the research is moving towards interactive systems, that take in both user feedback and automated recognition.
No comments:
Post a Comment