A Binocular, Foveated Active Vision System

This report documents the design and implementation of a binocular, foveated active vision system as part of the Cog project at the MIT Artificial Intelligence Laboratory. The active vision system features a three degree of freedom mechanical platform that supports four color cameras, a motion control system, and a parallel network of digital signal processors for image processing. To demonstrate the capabilities of the system, we present results from four sample visual-motor tasks.

Context-Based Vision System for Place and Object Recognition

While navigating in an environment, a vision system has to be able to recognize where it is and what the main objects in the scene are. In this paper we present a context-based vision system for place and object recognition. The goal is to identify familiar locations (e.g., office 610, conference room 941, Main Street), to categorize new environments (office, corridor, street) and to use that information to provide contextual priors for object recognition (e.g., table, chair, car, computer). We present a low-dimensional global image representation that provides relevant information for place recognition and categorization, and how such contextual information introduces strong priors that simplify object recognition. We have trained the system to recognize over 60 locations (indoors and outdoors) and to suggest the presence and locations of more than 20 different object types. The algorithm has been integrated into a mobile system that provides real-time feedback to the user.

A Comparison of Hardware Implementations for Low-Level Vision Algorithms

Early and intermediate vision algorithms, such as smoothing and discontinuity detection, are often implemented on general-purpose serial, and more recently, parallel computers. Special-purpose hardware implementations of low-level vision algorithms may be needed to achieve real-time processing. This memo reviews and analyzes some hardware implementations of low-level vision algorithms. Two types of hardware implementations are considered: the digital signal processing chips of Ruetz (and Broderson) and the analog VLSI circuits of Carver Mead. The advantages and disadvantages of these two approaches for producing a general, real-time vision system are considered.

Towards Autonomous Motion Vision

Earlier, we introduced a direct method called fixation for the recovery of shape and motion in the general case. The method uses neither feature correspondence nor optical flow. Instead, it directly employs the spatiotemporal gradients of image brightness. This work reports the experimental results of applying some of our fixation algorithms to a sequence of real images where the motion is a combination of translation and rotation. These results show that parameters such as the fization patch size have crucial effects on the estimation of some motion parameters. Some of the critical issues involved in the implementaion of our autonomous motion vision system are also discussed here. Among those are the criteria for automatic choice of an optimum size for the fixation patch, and an appropriate location for the fixation point which result in good estimates for important motion parameters. Finally, a calibration method is described for identifying the real location of the rotation axis in imaging systems.

The Three-Dimensional Interpretation of a Class of Simple Line-Drawings

We provide a theory of the three-dimensional interpretation of a class of line-drawings called p-images, which are interpreted by the human vision system as parallelepipeds ("boxes"). Despite their simplicity, p-images raise a number of interesting vision questions: *Why are p-images seen as three-dimensional objects? Why not just as flatimages? *What are the dimensions and pose of the perceived objects? *Why are some p-images interpreted as rectangular boxes, while others are seen as skewed, even though there is no obvious distinction between the images? *When p-images are rotated in three dimensions, why are the image-sequences perceived as distorting objects---even though structure-from-motion would predict that rigid objects would be seen? *Why are some three-dimensional parallelepipeds seen as radically different when viewed from different viewpoints? We show that these and related questions can be answered with the help of a single mathematical result and an associated perceptual principle. An interesting special case arises when there are right angles in the p-image. This case represents a singularity in the equations and is mystifying from the vision point of view. It would seem that (at least in this case) the vision system does not follow the ordinary rules of geometry but operates in accordance with other (and as yet unknown) principles.

Surface Reflectance Estimation and Natural Illumination Statistics

Humans recognize optical reflectance properties of surfaces such as metal, plastic, or paper from a single image without knowledge of illumination. We develop a machine vision system to perform similar recognition tasks automatically. Reflectance estimation under unknown, arbitrary illumination proves highly underconstrained due to the variety of potential illumination distributions and surface reflectance properties. We have found that the spatial structure of real-world illumination possesses some of the statistical regularities observed in the natural image statistics literature. A human or computer vision system may be able to exploit this prior information to determine the most likely surface reflectance given an observed image. We develop an algorithm for reflectance classification under unknown real-world illumination, which learns relationships between surface reflectance and certain features (statistics) computed from a single observed image. We also develop an automatic feature selection method.

Recognition of Surface Reflectance Properties from a Single Image under Unknown Real-World Illumination

This paper describes a machine vision system that classifies reflectance properties of surfaces such as metal, plastic, or paper, under unknown real-world illumination. We demonstrate performance of our algorithm for surfaces of arbitrary geometry. Reflectance estimation under arbitrary omnidirectional illumination proves highly underconstrained. Our reflectance estimation algorithm succeeds by learning relationships between surface reflectance and certain statistics computed from an observed image, which depend on statistical regularities in the spatial structure of real-world illumination. Although the algorithm assumes known geometry, its statistical nature makes it robust to inaccurate geometry estimates.

The Role of Fixation and Visual Attention in Object Recognition

This research project is a study of the role of fixation and visual attention in object recognition. In this project, we build an active vision system which can recognize a target object in a cluttered scene efficiently and reliably. Our system integrates visual cues like color and stereo to perform figure/ground separation, yielding candidate regions on which to focus attention. Within each image region, we use stereo to extract features that lie within a narrow disparity range about the fixation position. These selected features are then used as input to an alignment-style recognition system. We show that visual attention and fixation significantly reduce the complexity and the false identifications in model-based recognition using Alignment methods. We also demonstrate that stereo can be used effectively as a figure/ground separator without the need for accurate camera calibration.

Specialization of Perceptual Processes

In this report, I discuss the use of vision to support concrete, everyday activity. I will argue that a variety of interesting tasks can be solved using simple and inexpensive vision systems. I will provide a number of working examples in the form of a state-of-the-art mobile robot, Polly, which uses vision to give primitive tours of the seventh floor of the MIT AI Laboratory. By current standards, the robot has a broad behavioral repertoire and is both simple and inexpensive (the complete robot was built for less than $20,000 using commercial board-level components). The approach I will use will be to treat the structure of the agent's activity---its task and environment---as positive resources for the vision system designer. By performing a careful analysis of task and environment, the designer can determine a broad space of mechanisms which can perform the desired activity. My principal thesis is that for a broad range of activities, the space of applicable mechanisms will be broad enough to include a number mechanisms which are simple and economical. The simplest mechanisms that solve a given problem will typically be quite specialized to that problem. One thus worries that building simple vision systems will be require a great deal of {it ad-hoc} engineering that cannot be transferred to other problems. My second thesis is that specialized systems can be analyzed and understood in a principled manner, one that allows general lessons to be extracted from specialized systems. I will present a general approach to analyzing specialization through the use of transformations that provably improve performance. By demonstrating a sequence of transformations that derive a specialized system from a more general one, we can summarize the specialization of the former in a compact form that makes explicit the additional assumptions that it makes about its environment. The summary can be used to predict the performance of the system in novel environments. Individual transformations can be recycled in the design of future systems.

Feature Extraction Without Edge Detection

Information representation is a critical issue in machine vision. The representation strategy in the primitive stages of a vision system has enormous implications for the performance in subsequent stages. Existing feature extraction paradigms, like edge detection, provide sparse and unreliable representations of the image information. In this thesis, we propose a novel feature extraction paradigm. The features consist of salient, simple parts of regions bounded by zero-crossings. The features are dense, stable, and robust. The primary advantage of the features is that they have abstract geometric attributes pertaining to their size and shape. To demonstrate the utility of the feature extraction paradigm, we apply it to passive navigation. We argue that the paradigm is applicable to other early vision problems.

Hierarchical Object Recognition Using Libraries of Parameterized Model Sub-Parts

This thesis describes the development of a model-based vision system that exploits hierarchies of both object structure and object scale. The focus of the research is to use these hierarchies to achieve robust recognition based on effective organization and indexing schemes for model libraries. The goal of the system is to recognize parameterized instances of non-rigid model objects contained in a large knowledge base despite the presence of noise and occlusion. Robustness is achieved by developing a system that can recognize viewed objects that are scaled or mirror-image instances of the known models or that contain components sub-parts with different relative scaling, rotation, or translation than in models. The approach taken in this thesis is to develop an object shape representation that incorporates a component sub-part hierarchy- to allow for efficient and correct indexing into an automatically generated model library as well as for relative parameterization among sub-parts, and a scale hierarchy- to allow for a general to specific recognition procedure. After analysis of the issues and inherent tradeoffs in the recognition process, a system is implemented using a representation based on significant contour curvature changes and a recognition engine based on geometric constraints of feature properties. Examples of the system's performance are given, followed by an analysis of the results. In conclusion, the system's benefits and limitations are presented.

MARVEL: A System for Recognizing World Locations with Stereo Vision

To use a world model, a mobile robot must be able to determine its own position in the world. To support truly autonomous navigation, I present MARVEL, a system that builds and maintains its own models of world locations and uses these models to recognize its world position from stereo vision input. MARVEL is designed to be robust with respect to input errors and to respond to a gradually changing world by updating its world location models. I present results from real-world tests of the system that demonstrate its reliability. MARVEL fits into a world modeling system under development.

Vision by Man and Machine

The development of increasingly sophisticated and powerful computers in the last few decades has frequently stimulated comparisons between them and the human brain. Such comparisons will become more earnest as computers are applied more and more to tasks formerly associated with essentially human activities and capabilities. The expectation of a coming generation of "intelligent" computers and robots with sensory, motor and even "intellectual" skills comparable in quality to (and quantitatively surpassing) our own is becoming more widespread and is, I believe, leading to a new and potentially productive analytical science of "information processing". In no field has this new approach been so precisely formulated and so thoroughly exemplified as in the field of vision. As the dominant sensory modality of man, vision is one of the major keys to our mastery of the environment, to our understanding and control of the objects which surround us. If we wish to created robots capable of performing complex manipulative tasks in a changing environment, we must surely endow them with (among other things) adequate visual powers. How can we set about designing such flexible and adaptive robots? In designing them, can we make use of our rapidly growing knowledge of the human brain, and if so, how at the same time, can our experiences in designing artificial vision systems help us to understand how the brain analyzes visual information?

Vision, Instruction, and Action

This thesis describes Sonja, a system which uses instructions in the course of visually-guided activity. The thesis explores an integration of research in vision, activity, and natural language pragmatics. Sonja's visual system demonstrates the use of several intermediate visual processes, particularly visual search and routines, previously proposed on psychophysical grounds. The computations Sonja performs are compatible with the constraints imposed by neuroscientifically plausible hardware. Although Sonja can operate autonomously, it can also make flexible use of instructions provided by a human advisor. The system grounds its understanding of these instructions in perception and action.

Teaching an Old Robot New Tricks: Learning Novel Tasks via Interaction with People and Things

As AI has begun to reach out beyond its symbolic, objectivist roots into the embodied, experientialist realm, many projects are exploring different aspects of creating machines which interact with and respond to the world as humans do. Techniques for visual processing, object recognition, emotional response, gesture production and recognition, etc., are necessary components of a complete humanoid robot. However, most projects invariably concentrate on developing a few of these individual components, neglecting the issue of how all of these pieces would eventually fit together. The focus of the work in this dissertation is on creating a framework into which such specific competencies can be embedded, in a way that they can interact with each other and build layers of new functionality. To be of any practical value, such a framework must satisfy the real-world constraints of functioning in real-time with noisy sensors and actuators. The humanoid robot Cog provides an unapologetically adequate platform from which to take on such a challenge. This work makes three contributions to embodied AI. First, it offers a general-purpose architecture for developing behavior-based systems distributed over networks of PC's. Second, it provides a motor-control system that simulates several biological features which impact the development of motor behavior. Third, it develops a framework for a system which enables a robot to learn new behaviors via interacting with itself and the outside world. A few basic functional modules are built into this framework, enough to demonstrate the robot learning some very simple behaviors taught by a human trainer. A primary motivation for this project is the notion that it is practically impossible to build an "intelligent" machine unless it is designed partly to build itself. This work is a proof-of-concept of such an approach to integrating multiple perceptual and motor systems into a complete learning agent.