Monocular Depth Estimation by Non-Local Decoder-Squeeze-and-Excitation Network with Adaptive Depth List

Monocular depth estimation is an important topic in computer vision. Recently the CNNs (Convolutional Neural Networks) based model shows a reasonable result from an end-to-end encoder-decoder architecture. For the encoder part, most of the research is based on a robust feature extractor to get good features. With a strong encoder, it was found that even simple up-sampling processes can achieve good accuracy. However, the decoder part is more critical in a high-quality depth estimation task. Even now, there is no intuitive way to calibrate the feature map for the upsampling process. In this paper, we present a novel monocular depth estimation design. We propose an innovative CNN-based network module that considers the whole up-sampling process globally. This design is based on the concept of SE-Net, and properly recalibrated the feature maps with a global perspective attention mechanism. We further combine it with Non-local network attention mechanisms to design the Non-Local Decoder-Squeeze-and-Excitation (NL-DSE) module for the whole up-sampling process. Furthermore, we also propose an output limiting range method called Adaptive Depth List (ADL) to enhance the precision of the near-distance estimation. Combining these proposed techniques, our results are evaluated on the NYU Depth V2 dataset and outperform the state-of-the-art CNN-based approaches in accuracy.