UNSUPERVISED MONOCULAR DEPTH ESTIMATION OF DRIVING SCENES USING SIAMESE CONVOLUTIONAL LSTM NETWORKS

摘要

Estimating depth from a single RGB image is an active research topic in computer vision because of its broad applications in scene understanding, autonomous driving, and traffic surveillance systems. This task involves estimating a pixel-wise depth map from a single image. Significant progress has been made on monocular depth estimation using deep learning-based techniques. Current approaches employ geometry-based image reconstruction methods instead of ground truth depth labels to perform depth estimation in an unsupervised manner. In this paper, we present a deep learning model to simultaneously learn and refine depth maps from a single RGB image and in an end-to-end manner by casting the monocular depth estimation as an image reconstruction problem. We propose an unsupervised framework for monocular depth estimation that trains a Siamese convolutional long short-term memory (Siamese convLSTM) network to jointly perform estimation and refinement of depth maps using rectified stereo image pairs and produce a depth map from a single RGB image at test time. Experimental results show that simultaneously performing these two tasks leads to improving depth estimation accuracy. In particular, using the KITTI 2015 driving dataset for evaluation, our proposed Siamese convLSTM network achieves excellent performance on monocular depth estimation, both quantitatively and qualitatively.