Programming Language Support for Digitized Images or, The Monsters in the Closet

Computer vision (image understanding) algorithms are difficult to write, debug, maintain, and share. This complicates collaboration, teaching, and replication of research results. This paper shows how user-level code can be simplified by providing better programming language constructs, particularly a new abstract data type called a "sheet." These primitives have been implemented as an extension to Scheme. Implementation of sheet operations is made challenging by the fact that images are extremely large, e.g. sometimes over 5 megabytes each. Therefore, operations that loop through images must be compiled from (a specialized subset of) Scheme into C. This paper discusses how the need for extreme efficiency affects the design of the user-level language, the run-time support, and the compiler.