Depth Estimation from Single RGB Image using Attention based Encoder-Decoder Network

Can we estimate depth from a single RGB Image? Depth plays a crucial part in understanding the 3D structure of a scene. The task becomes more complex if only one image of the scene is at the disposal. This research paper proposes a method to find depth from a single RGB Image using Attention based Encoder-Decoder convolution architecture. EfficientNetB7 is used as the encoder backbone whereas Bi-linear interpolation is used as the decoder for the network. Apart from the encoder-decoder architecture, we have also utilized the skip-connections and attention blocks to focus on the important features for depth estimation. Our results on NYUDepthV2 benchmark dataset demonstrates the effectiveness of the model and it’s potential for applications like Robotics and Augmented Reality. We provide a detailed qualitative and quantitative analysis of our proposed method.