Joint Bilateral Filter and Multi-Scale Cost Aggregation in Stereo Matching

Dense correspondence is a key problem in binocular stereo vision. The existing solution for this problem can be divided into local method and global or semi-global method. Bilaterally weighted patches matching is a classical local method, while it is computationally expensive and its accuracy need to be improved. In this paper, a new method is proposed which uses bilaterally weighted method to calculate matching cost and multi-scale method to produce cost aggregation. Use Gauss down-sampling method to produce multi-scale image, calculate matching cost in every sample scale. Then add inter-scale regularization into optimization function and solve the new optimization problem. This new algorithm is evaluated on Middlebury dataset, and it presents significant improvement than single bilaterally weighted method. Introduction Stereo matching algorithm generally comprises four steps: cost computation, cost aggregation, disparity computation, disparity refinement [1]. In cost computation, the matching cost of each pixel will be calculated, and then, the costs are aggregated, finally, the disparity of each pixel is calculated by different methods and the disparity will be refined by some post-processing methods. In every steps, most methods have been proposed, different methods have different impact on stereo matching algorithm. Actually, the choice of cost aggregation methods has most significant impact on the performance of matching algorithm. Some simple linear filters all can be used to produce cost aggregation such as box and Gaussian filter. Yoon and Kweon [2] use bilateral filter method to compute cost aggregation. Rhemann et al. [3] adopted the guided image filter into coast aggregation. Yang [4] proposed a non-local method in recent which is different from the local method and its kernel size is the entire image. So far, all the algorithms are only processing the input images at the normal scale. Considering that eye processing information is a process from rough to precision, so we proposed to compute the cost aggregation in multiple scale. In this paper, our algorithm aim to improve the performance of the bilateral filter method. The algorithm can be summarized in following four parts: 1) compute sampling images at different scale 2) compute match cost and aggregate cost using bilateral filter kernel 3) calculate robust cost aggregation combing cost aggregation computed in step 2) 4) produce the disparity and refine the disparity The method The method can be divided into four parts as introduced above. Firstly, use Gauss down-sampling method to produce the multi-scale images. Then calculate the matching cost of every pixel. And next, aggregate the matching cost. The most important part of our method is using the multi-scale cost aggregation model and inter-scale regularization model to calculate the matching cost of every pixel. Finally, select the lowest cost as the matching cost of each pixel. The frame of our algorithm can be described as follows: 4th International Conference on Mechatronics, Materials, Chemistry and Computer Engineering (ICMMCCE 2015) © 2015. The authors Published by Atlantis Press 2945