The paper by D. Marr and H. K. Nishihara explores the computational problems associated with deriving useful information from retinal images for the purpose of representing and recognizing three-dimensional shapes. The authors present three criteria—accessibility, scope & uniqueness, and stability & sensitivity—to judge the usefulness of a shape representation for recognition. They consider three aspects of a representation's design: its coordinate system, primitives, and organization. A shape representation for recognition should use an object-centered coordinate system, include volumetric primitives of varied sizes, and have a modular organization. The basic process for deriving a shape description involves identifying the natural axes of a shape in its image and transforming viewer-centered axis specifications to an object-centered coordinate system. Shape recognition involves a collection of stored shape descriptions and various indexes to associate a newly derived description with an appropriate stored description. The authors describe a conservative recognition process that can extract more information from the image by relaxing constraints. The paper also discusses the advantages of a modular organization in shape descriptions, the coordinate system of the 3-D model representation, and the interaction between the derivation and recognition processes.The paper by D. Marr and H. K. Nishihara explores the computational problems associated with deriving useful information from retinal images for the purpose of representing and recognizing three-dimensional shapes. The authors present three criteria—accessibility, scope & uniqueness, and stability & sensitivity—to judge the usefulness of a shape representation for recognition. They consider three aspects of a representation's design: its coordinate system, primitives, and organization. A shape representation for recognition should use an object-centered coordinate system, include volumetric primitives of varied sizes, and have a modular organization. The basic process for deriving a shape description involves identifying the natural axes of a shape in its image and transforming viewer-centered axis specifications to an object-centered coordinate system. Shape recognition involves a collection of stored shape descriptions and various indexes to associate a newly derived description with an appropriate stored description. The authors describe a conservative recognition process that can extract more information from the image by relaxing constraints. The paper also discusses the advantages of a modular organization in shape descriptions, the coordinate system of the 3-D model representation, and the interaction between the derivation and recognition processes.