Self-supervised Monocular Depth Estimation with Uncertainty-aware Feature Enhancement and Depth Fusion

Due to insufficient model training and other reasons, the neural network may not be able to accurately fit the implicit mapping relationship between scene depth and 2D images, resulting in the network's uncertain output. In the monocular depth prediction task, the uncertainty based on the predicted depth represents the image understanding deviation of the network and the confidence of the predicted depth. Therefore, we design the uncertainty-aware feature enhancement module (UFE) and the uncertainty-guided multi-scale depth fusion module (UMDF); besides, combined with Nlonodepth2, a self-supervised uncertainty -aware general framework for monocular depth prediction is proposed, uncertainty -aware feature enhancement modules are embedded in the decoder to strengthen the image understanding ability of the model. In addition, we use an uncertainty -guided multi -scale depth fusion module to optimize the coarse depth predicted by the network adaptively. The abundant experiments conducted on KITTI and Cityscapes certify that the two uncertainty-based modules can dramatically boost the model's performance and flexibly he embedded into different self-supervised models.