Multi-thread Solution of Permutohedral Refined UNet for Cloud/Shadow Detection in High-resolution Remote Sensing Images

Boundary-aware high-resolution segmentation aims to partition a high-resolution image into regions in terms of both semantic information and low-level visual features, which has been applied to the discovery of fine-grained objects of interest, for example, cloud/shadow detection in remote sensing images. Their computational cost, on the other hand, has to be carefully considered due to the quadratic time complexity of their naive implementations. Such limitations to practical applications motivate us to try to improve the efficiency performance from a practical perspective by distributing independent computations into multiple CPU threads. We therefore present a multi-thread implementation for our Permutohedral Refined UNet to achieve global boundary refinement for cloud/shadow detection. Specifically, the bilateral/spatial feature generations, a part of filter initialization, a part of filter computation, and CRF iterations can be computed in parallel, which allows us to distribute such computations into multiple CPU threads. The left computations still run sequentially. We then evaluate the efficiency performance of our multi-thread implementations in statistics, and find that a significant efficiency gain is achieved by our multi-thread implementations. The source codes are publicly available at https://github.com/jiaolobel/perm-refined-unet-efficient-impls.