Self-Supervised Learning of Depth From Sequence of Images

The Estimation of distance between objects in a relation to camera is still a challenging problem of many robotics applications, such as self-driving cars, 3D scene reconstruction, and robot grasping. Estimation of depth at which different objects are present in the scene are previously approached using many supervised based learning methods. Our Self-supervised learning based architecture has demonstrated great potential in estimating depth from Monocular images. SFM (Structure from Motion) and depth animation based on stereo images, basically rely on the feature matching of different views of the same scene. Capturing depth related data from a single photograph would be a challenging task (ill-posed problem). In this work, We tried to present a self-supervised learning based monocular depth estimation method. Further, we also addressed some of the common issues caused because of dynamic objects and occlusions. We have evaluated the results of the monocular sequence based method with stereo image based method and we observed that with no additional contribution the model performs poorly and addition of masking, min re-projection loss improves the performance of the model.