Semi-Supervised Adversarial Monocular Depth Estimation

In this paper,we address the problem of monocular depth estimation with limited number of training depth maps.With the estimated depth value,many tasks in computer vision can get a performance improvement compared with only using RGB images.This includes reconstruction,recognition,semantic segmentation,and human pose estimation.By combining depth information,these tasks can better distinguish the boundaries and relative positions of objects,and thus better complete the corresponding tasks.As for monocular depth estimation,most existing methods rely on an encoder-decoder CNN with a meticulously designed loss function to regress depth values.However,such methods suffer edge blurring and require a vast number of image-depth pairs(or stereo image pairs)to get satisfactory performance.In this paper,we propose a semi-supervised adversarial framework,which needs just a small number of image-depth pairs along with a large amount of cheaply-available monocular images to learn the mapping function between the image and the depth.In particular,we design a generator that has both long skip connections and short skip connections to regress depth and two discriminators to evaluate the predicted depth without looking at the corresponding depth.This is achieved by passing feedbacks of two discriminators(i.e.one distinguishes the predicted depth from the real depth,and the other judging whether the image and predicted depth pair comes from the real joint distribution)to the generator network as a loss.We have demonstrated that,with an order of magnitude less image-depth training pair than that of the traditional methods with full supervision,our method can better preserve the object edge in depth map and outperform the state-of-the-arts.