Lightweight Monocular Depth Estimation Based On Perceptual Loss and Network Slimming

Seungjae Lee, Yongsik Lee, Suwoong Lee, Jong-Gook Ko

Abstract: The performance of deep learning-based depth estimation depends on encoding layers, decoding layers, and loss function. In this paper, we propose a lightweight monocular depth estimation method by considering those three elements. First, we introduce perceptual loss to represent the characteristic of the depth information and the proposed loss improves accuracy. Second, encoding and decoding layers are analyzed and redesigned for lightweight depth estimation. Finally, encoding layers are pruned for more efficient architecture. In the experiment, the proposed method is evaluated on the NYU Depth V2 dataset and KITTI dataset and it shows that proposed loss and network slimming improve the accuracy of depth estimation with less memory and computational complexity. The proposed model is also deployed to the mobile device, iPhone 11 Pro. It runs more this 30 fps and the inference time is 5 ms.