12 June 2018 in IEEE Access
3-D vision plays an important role in intelligent perception of robot, while it requires extra 3-D sensors. Depth estimation from monocular videos provides an alternative mechanism to recover the 3-D information. In this paper, we propose an unsupervised learning framework that uses the perceptual loss for depth estimation. Depth and pose networks are first trained to estimate the depth and the camera motion of the video sequence, respectively. With the estimated depth and pose of the original frame, the adjacent frame can be reconstructed. The pixel-wise differences between the constructed frame and the original frame are used as per-pixel loss. Meanwhile, reconstructed views and original views can be used to extract advanced features from a pre-trained network to define and optimize perceptual loss functions to assess the quality of reconstructions. We combine the respective advantages of these two methods and present an approach of generating a depth map by training the feed-forward network with per-pixel loss function and perceptual loss function. The experimental results show that our method can significantly improve the estimation accuracy of depth map.
28 January 2020 in IEEE Transactions on Image Processing
In recent years, supervised deep learning methods have shown a great promise in dense depth estimation. However, massive high-quality training data are expensive and impractical to acquire. Alternatively, self-supervised learning-based depth estimators can learn the latent transformation from monocular or binocular video sequences by minimizing the photometric warp error between consecutive frames, but they suffer from the scale ambiguity problem or have difficulty in estimating precise pose changes between frames. In this paper, we propose a joint self-supervised deep learning pipeline for depth and ego-motion estimation by employing the advantages of adversarial learning and joint optimization with spatial-temporal geometrical constraints. The stereo reconstruction error provides the spatial geometric constraint to estimate the absolute scale depth. Meanwhile, the depth map with an absolute scale and a pre-trained pose network serves as a good starting point for direct visual odometry (DVO). DVO optimization based on spatial geometric constraints can result in a fine-grained ego-motion estimation with the additional backpropagation signals provided to the depth estimation network. Finally, the spatial and temporal domain-based reconstructed views are concatenated, and the iterative coupling optimization process is implemented in combination with the adversarial learning for accurate depth and precise ego-motion estimation. The experimental results show superior performance compared with state-of-the-art methods for monocular depth and ego-motion estimation on the KITTI dataset and a great generalization ability of the proposed approach.
19 June 2019 in IEEE Access