Depth Map Preparation and Salient Object Segmentation using Focal stack

Abstract

This thesis presents a method for generating depth maps from a focal stack of images and utilizing the depth information for salient object detection and binary segmentation. The focal stack dataset used is the Mobile Depth dataset captured using a mobile phone camera. The approach involves aligning the frames in the focal stack, computing a sharpness map using a discrete cosine transform (DCT) based focus measure, refining the sharpness map through edge-preserving filtering, and estimating the depth map by a weighted combination of frame indices. The depth map is then employed as a saliency map for salient object detection. An adaptive thresholding technique based on Otsu’s method generates a trimap, which is fed into the GrabCut algorithm to produce a high-quality binary segmentation mask of the salient object. Challenges addressed include handling textureless regions, achieving accurate depth estimation with limited sampling frequency, and preserving edge details during filtering. The proposed method aims to leverage depth information from focal stacks to enhance salient object detection and segmentation performance, with potential applications in areas such as computer vision and image processing.