cdcseacave
Is it suitable to achieve DensifyPointClould with CUDA?
I want to use GPU programming to optimize the running time of the Reconstruction of DensePointCloud. May I ask if it's feasible to rewrite some functions in SceneDensify with CUDA to achieve that? If possible, could I directly use some kernel functions you wrote in the SceneRefineCUDA to reduce the amout of work? Could you give me some suggestions? Thanks in advance!
yes, the patchmatch algorithm greatly benefits from a GPU implementation and for that EstimateDepthMapTmp() should be ported; however the code can not just be directly ported to CUDA, the propagation step (the way the depth-map image is parsed) needs to be changed from pseudo serial mode as it is now, to multithreded mode as required by a GPU; this can be achieved by using a checkerboard approach, however the propagation can suffer; this is well explained in Xu and Tao - 2019 - Multi-Scale Geometric Consistency Guided Multi-Vie
