Depth Prediction from Monocular Images with CGAN

Depth prediction from monocular images is an important task in many computer vision fields as monocular cameras are currently the majorities of the image acquisition equipment, which is used in many fields such as stereo scenes understanding and Simultaneous Location and Mapping (SLAM). In this paper, we regard depth prediction as an image generation task and propose a new method for monocular depth prediction using Conditional Generative Adversarial Nets (CGAN). We transform the corresponding depth images of RGB images as the Relative depth images by dividing the maximum value, then we use an encoder-decoder as the generator of CGAN, which is used to generate depth images corresponding to input RGB images, the discriminator is constituted by an encoder, which is used to discriminate whether the input images are true or fake by evaluating the difference between input images. By learning the potential correspondence between pixels of RGB images and depth image, we could finally obtain the corresponding depth images of test RGB images with our CGAN model. We test our model with different objective functions in TUM RGB-D dataset and NYU V2 dataset, and the result shows excellent performance.