Estimating Depth Map of an RGB image using Encoders and Decoders

Abstract

Depth estimation from a single RGB picture has emerged as one of the most significant study areas in recent years because of the wide variety of applications, from robotics to medical sciences. Monocular depth estimation has often had low resolution and blurry depth maps, which are not usable for further training of models with specific applications. The main drawback of generic depth estimation models is that they take an object and its environment into consideration. Because of this, traditional deep learning-based systems often experience severe setbacks in forecasting depths. This paper proposes an encoder-decoder network that, using transfer learning, can forecast high-quality depth pictures from a single RGB image. After initialising the encoder using augmentation algorithms and significant feature extraction from pre-trained networks, the decoder predicts the high-end depth maps. We have also used several boundary detection techniques to remove the object from its environment without losing the object's pixel information. Our network performs comparable to the state-of-the-art on two datasets and also generates qualitatively better results that more accurately represent object boundaries which can be further used in 6D pose estimation to perform robotic grasping.