EvisionNet:An Unsupervised Learning Method For Intrinsic,Ego-Motion And Scene Depth

Scene depth,ego-motion and intrinsic are important for robot navigation,industrial control and autonomous driving.With the development of artificial intelligence and the establishment of large open datasets,sense depth prediction has been very successful with supervised and unsupervised deep neural networks.Despite the high performance of these methods,most of them,including the unsupervised ones,require intrinsic as input,which needs camera calibration.In this work we address EvisionNet,an unsupervised learning method of scene depth,egomotion and intrinsic.The functionality of intrinsic prediction allowing us to extract accurate depth and motion from arbitrary videos of unknown origin without camera calibration.Where input data is provided by monocular videos,as cameras are the cheapest,least restrictive and most ubiquitous sensor for robotics.The EvisionNet proposed in this paper mainly consists of two subnetworks of encoderdecoder structure: DepthNet for scene depth prediction and MotionNet for output the confidence mask.The mask is used to suppress the influence of untrusted pixels caused by scene motion,occlusion and other factors in the calculation of loss function.The subnetworks for ego-motion and intrinsic prediction shares the encoder with MotionNet.All models are trained synchronously by the joint loss function we proposed,which based on the reprojection loss for unsupervised learning.We give the necessary conditions to ensure the effectiveness.Empirical evaluation on the open datasets demonstrates the effectiveness of our approach.The Scene depth performs comparably with supervised methods,and ego-motion estimation performs favorably compared to established systems under comparable input settings.In addition,we demonstrate that intrinsic can be learned from monocular image sequence.