Supervised Estimation of Dense Optical Flow

Abstract

End-to-end trained Convolutional Neural Network (CNN) have signi cantly advanced the eld of computer vision in recent years, particularly high-level vision problems, because of its strong non-linear tting ability. In context of optical ow, obtaining dense, ground truth per-pixel for real scenes is di cult and thus rarely available. But CNN in recent years demonstrated that dense optical ow estimation can be cast as a learning problem. However, the state of the art with regard to the quality of the ow has still been de ned by traditional methods. In this thesis, rstly, we used a compact but e ective CNN model, called U-Net, which contains an encoder part and a decoder part and used benchmark datasets: MPI-Sintel, KITTI and Middlebury; for training and evaluation, in a supervised manner. Secondly, we used some traditional energy- based loss function for dense optical ow estimation. Thirdly, we used backward warping with bilinear interpolation to predict rst image and build occlusion mask using ground truth ow. Experimental results show that our proposed method is at par with state-of-the-art supervised CNN methods.