Bibliography
Tracy Hammond and Randall Davis. LADDER, A Sketching Language for User Interface Developers. Computers & Graphics. 29. 2005.
Summary
This paper discusses LADDER - a sketching language for user interface developers. The goal of the authors was to build a domain independent framework for developing sketch recognition interfaces, which can be used by developers to build a recognizer in their domain. The framework requires the developer to write a domain description, which is automatically translated into shape recognizers, editing recognizers and shape exhibitors, for the specific domain.
Shapes can be defined in ladder using primitive shapes or previously defined shapes. The definition of a shape consists of components (elements from which shape is built), constraints (geometric constraints), aliases, editing behavior (like 'cross on arrow' for erase) and display methods (what to display when object is recognized). This description is translated into shape recognizers, exhibitors and editors that can be used in the domain sketch interface. The system makes it easy to define these shapes by providing abstract definitions for reusability and hierarchical definitions for simplicity. Shapes can be grouped together using 'shape groups' to provide chaining effects while editing. LADDER also supports 'Vectors' for defining shapes with arbitrary components. (such as polylines)
One limitation of the system is that the user defined shapes must be solely constrained of using primitive constraints defined in ladder. This makes it difficult to define shapes that have a lot of irregularities. Ladder supports a vase array of constraints such as acute, obtuse, concentric, larger, near, slope etc. Some constraints are defined for a given orientation. It also allows users to specify rotational invariance to shapes.
LADDER's correctness in identifying shapes depends a lot on getting the primitive shapes right. This is because, once the primitive shapes have been identified (can be many), these are then used by a higher level recognizer (written as jess rules) that matches the data with all possible shape combinations that satisfy the rule. The system may select the wrong higher level shape, as a greedy approach is used for picking the higher level shape. (This is done to maintain efficiency and real-time editing capability) A drawing gesture is short-circuited whenever an editing gesture is matched. Finally, a shape constraint solver is used to display a shapes ideal strokes.
LADDER domain desriptions have been tested in a variety of domains such as mechanical engineering, UML class diagrams, finite state machines and flowcharts.
Discussion
LADDER is the first paper I am reading, which offers a full blown generic sketch recognition system since GRANDMA. (explained in Rubine's paper) I particularly like the abstractions defined in LADDER, that makes it look very modular and reusable. I wonder if ladder can be improved using feedback from past decisions made by it. Some kind of reinforcement learning. It will be interesting to combine such a system with other recognition techniques, that may not be geometric.
Tracy Hammond and Randall Davis. LADDER, A Sketching Language for User Interface Developers. Computers & Graphics. 29. 2005.
Summary
This paper discusses LADDER - a sketching language for user interface developers. The goal of the authors was to build a domain independent framework for developing sketch recognition interfaces, which can be used by developers to build a recognizer in their domain. The framework requires the developer to write a domain description, which is automatically translated into shape recognizers, editing recognizers and shape exhibitors, for the specific domain.
Shapes can be defined in ladder using primitive shapes or previously defined shapes. The definition of a shape consists of components (elements from which shape is built), constraints (geometric constraints), aliases, editing behavior (like 'cross on arrow' for erase) and display methods (what to display when object is recognized). This description is translated into shape recognizers, exhibitors and editors that can be used in the domain sketch interface. The system makes it easy to define these shapes by providing abstract definitions for reusability and hierarchical definitions for simplicity. Shapes can be grouped together using 'shape groups' to provide chaining effects while editing. LADDER also supports 'Vectors' for defining shapes with arbitrary components. (such as polylines)
One limitation of the system is that the user defined shapes must be solely constrained of using primitive constraints defined in ladder. This makes it difficult to define shapes that have a lot of irregularities. Ladder supports a vase array of constraints such as acute, obtuse, concentric, larger, near, slope etc. Some constraints are defined for a given orientation. It also allows users to specify rotational invariance to shapes.
LADDER's correctness in identifying shapes depends a lot on getting the primitive shapes right. This is because, once the primitive shapes have been identified (can be many), these are then used by a higher level recognizer (written as jess rules) that matches the data with all possible shape combinations that satisfy the rule. The system may select the wrong higher level shape, as a greedy approach is used for picking the higher level shape. (This is done to maintain efficiency and real-time editing capability) A drawing gesture is short-circuited whenever an editing gesture is matched. Finally, a shape constraint solver is used to display a shapes ideal strokes.
LADDER domain desriptions have been tested in a variety of domains such as mechanical engineering, UML class diagrams, finite state machines and flowcharts.
Discussion
LADDER is the first paper I am reading, which offers a full blown generic sketch recognition system since GRANDMA. (explained in Rubine's paper) I particularly like the abstractions defined in LADDER, that makes it look very modular and reusable. I wonder if ladder can be improved using feedback from past decisions made by it. Some kind of reinforcement learning. It will be interesting to combine such a system with other recognition techniques, that may not be geometric.
No comments:
Post a Comment