MFCS-Depth: An Economical Self-Supervised Monocular Depth Estimation Based on Multi-Scale Fusion and Channel Separation Attention

Self-supervised monocular depth estimation plays an extremely important role in fields such as autonomous driving and intelligent robot navigation. However, general monocular depth estimation models require massive computing resources, which seriously hinders their deployment on mobile devices, which is urgently needed in fields such as autonomous driving. To address this problem, we propose MFCS-Depth, an economical monocular depth estimation method based on multi-scale fusion and channel separation attention mechanism. We use the Transformer architecture with linear self-attention as its encoder to ensure its global modeling and economy. A high-performance and low-cost decoder has also been designed to improve the local and global reasoning of the network through multi-scale attention fusion and uses scale-wise channel separation to reduce parameters and computing costs significantly. Extensive experiments show that MFCS-Depth achieves competitive results with very few parameters on the KITTI and DDAD datasets and achieves state-of-the-art performance among methods of similar size.