PAR-mono: monocular video depth estimation network based on channel separation and dynamic attention

The current video depth estimation methodology is hindered by two significant limitations: inter-frame depth discontinuity and slow model convergence. To address these issues, we propose a novel depth estimation framework, PAR-mono, which treats video depth estimation as a cyclic regression problem. To address the problem of discontinuous depth estimation between model frames, we propose to use the depth estimation of the previous frame to assist the model’s depth estimation for the current frame. At the same time, we propose a feature-adaptive encoder, which utilises variability convolution to improve the spatial consistency of the depth map. Finally, in order to reduce In order to reduce the parameters and complexity of the SOTA model, we propose the Channel separation shuffle splicing module, which maintains the accuracy of the depth estimation as much as possible while improving the inference speed of the model. Furthermore, in order to refine the multi-scale information extracted by the encoder and enhance the extraction of global information, we propose the dynamic attention Transformer module. The efficacy of our method was evaluated on the KITTI datasets, demonstrating a notable enhancement in consistency, accuracy, and efficiency compared to previous approaches.