Unsupervised Depth Completion Guided by Semantic Edges

Currently, unsupervised depth completion techniques often encounter challenges in terms of restoration accuracy and detail preservation when dealing with images containing complex structures and fine textures. To this end, an unsupervised depth completion model guided by semantic segmentation edges was proposed in this paper. By introducing a semantic edge branch, the edge information of semantic segmentation features was utilized to precisely delineate object contours, guiding the network to generate more natural and coherent restoration results. Through a dual-attention feature fusion module, depth features and RGB image features were effectively combined, significantly enhancing the model’s capability to extract and learn structural features. On the KITTI dataset, the RMSE index decreases by approximately 2.00% compared with other models, and on the VOID dataset, the RMSE index decreases by approximately 9.18% compared with other models, indicating that the model is effective.