Quantifying the Contextual Separability of Visual Confusion

Visual information and contextual knowledge can both contribute to the overall performance of an object recognition system. However, little work has so far been done to quantify the extent to which context can compensate for visual error and confusion. Although contextual patterns are essentially fixed, visual confusion is largely determined by visual features, and quantifying the ´compatibility´ of context and confusion can help estimate how much visual performance might improve or even deteriorate with the application of contextual knowledge. Powerful contextual cues might allow the use of cheaper and less discriminative visual processing to achieve an acceptable level of performance in combination, as might biasing the training of a visual system towards classes that are contextually difficult to separate. Knowledge of incompatibility would necessitate a different choice of visual features. This paper develops a measure of ´compatibility´ that attempts to determine whether the particular patterns of confusion in a given visual system can be separated by context. A reclassification method of utilising context to improve basic visual classification is given and experiments testing whether the measure does predict when context can separate visual confusion are presented. The results for the compatibility measure and reclassification method are promising.