Illumination invariant object recognition using the MNS method

Abstract

The suitability of the Multimodal Neighbourhood Signature (MNS) method for illumination invariant recognition is investigated. The MNS algorithm directly formulates the problem of extracting illumination invariants from local colour appearance of an object. The invariants are the channel-wise ratio and the cross-ratio computed from modes (pairs of modes respectively) of colour density function in neighbourhoods with multimodal density function. The MNS algorithm is tested on a colour object recognition task designed to test the effectiveness of algorithms claiming illumination invariance properties. The image set used is publicly available from the Simon Fraser University. Results previously reported using colour constancy and histogram matching were comparable to the performance of the presented method that achieved recognition rate of 60%. When the pose of the objects was fixed recognition performance was 84%.