This is my final project of Computer Vision: from recognition to geometry course. This project is held by Mediatek.
In this project, we were given ten pairs of realistic stereo images, all of them are from the custom of Mediatek. For these images, depth estimation is challenging since the presence of negative impact such as reflection and bokeh effect. To address the issue, we designed a specific pipelined algorithm. Besides optimizing the existing parts, we add following extensions to further improve the accuracy and reliability: histogram equalization, image segmentation, and using SURF to estimate the range of disparity. We won the fist prize in the competitoin of this project because of our innovation and performance.