Vehicle Retrieval Using Eigen Color and Multiple Instance Learning

This paper presents a novel approach for retrieving images from databases using eigen color and the concept of multiple instance learning. Usually, vehicles have various colors and shapes under different viewpoints, weathers, and lighting conditions. All the variations will increase many difficulties and challenges in selecting a general feature to describe vehicles. Thus, traditional methods to retrieve vehicles require their orientations or colors being fixed. To tackle this problem, this paper proposes a novel vehicle retrieval system for effectively retrieving vehicles from databases no matter what orientations and colors they are. First of all, this paper proposes a novel color transform model, which is global and does not need to be re-estimated for any new vehicles or new images, to extract different regions of interest (or vehicle analogues) from databases. Then, to more accurately locate desired vehicle images, this paper uses the MIL (multiple-instance learning) method to learn specific visual properties of vehicles from query images. However, the MIL technique requires the positive training data being strongly positive and the negative ones being strongly negative. This requirement is too constrained in real cases and will lead to lots of false detection. This problem can be easily tackled if an eigen color transform is introduced. The extra consideration "eigen color" will add more capabilities to the MIL learner for capturing the embedded concept more accurately. Furthermore, during the learning process, since no time-consuming optimization process is involved, all the desired visual concept can be obtained immediately and adapted to different user's requests. Experimental results reveal the feasibility and high accuracy of the proposed approach in vehicle retrieval system.