Uncertainty Propagation in Shape Reconstruction and Moving Object Detection from Optical Flow

Abstract

Computer vision systems often have to deal with information that is inherently uncertain. Consider, for example, an object detection system by vision. It usually consists of several modules which execute their specific tasks. Low-level modules in the system detect image primitives such as edges and textures. However, it is hopeless to expect those modules to perfectly achieve their tasks. One reason is the existence of noise. Noise contaminates input images and causes uncertainty to the results. The other reason is that image primitives have inherent uncertainty. Edges vary in strength from very weak ones to obvious ones, so the edge detection module can only output a fuzzy result. This situation is the same in the texture analyzing module. High-level modules, which are the object detection modules in the system, must make a decision based on information with uncertainty about the image primitives. Therefore, it is very important that these modules are so designed that they can handle the uncertainty and also propagate information about the uncertainty from the inputs to outputs.