Shape completion with graph attention layer and surface normal
Shape completion, the problem of inferring the complete geometry of an object given the partial observation, is an essential part of robotics and computer vision. This work proposes a novel graph neural network, Graph Normal Network (GNnet). Its encoder combines a graph-based model for encoding loca...
| Published: |
|
|---|---|
| Online Access: | http://hdl.handle.net/2047/D20409245 |
| Summary: | Shape completion, the problem of inferring the complete geometry of an object given the partial observation, is an essential part of robotics and computer vision. This work proposes a novel graph neural network, Graph Normal Network (GNnet). Its encoder combines a graph-based model for encoding local information with an MLPs-based architecture for encoding global information. Its decoder features the usage of the surface normal to sample neighbors to densify the output precisely. Our experiments show that GNnet could generate dense, compelling completions from partial observations and significantly outperforms state-of-the-art completion methods on the Shapenet dataset.--Author's abstract |
|---|