Dense Correspondence Learning between Key Points in 3D C.elegans Based on Depth Map Matching

The positioning and type identification of 3D C.elegans cells are key steps in establishing a cell atlas. Therefore, this paper introduces a learnable end-to-end framework-based on the depth map matching for three-dimensional C.elegans type recognition, utilized to create a strong correlation among the flexible point clouds of C.elegans. We use graph matching to solve correspondence estimates and use depth map matching to better exploit local geometric information and global structural information. We specifically suggest the creation of a graph generator that takes multiple layers of multidimensional features as nodes and utilizes attention and co-attention mechanisms in transformers to build the graph while calculating point-wise Euclidean distance. The structure and topological geometry information of the graph is embedded into the nodes through the graph-based feature extractor and the final estimation of the matching matrix linking the two point clouds has been completed by Affinity Layer and Sinkhorn Layer.