Dual octree graph networks for learning adaptive volumetric shape representations

We present an adaptive deep representation of volumetric fields 3D shapes and efficient approach to learn this for high-quality shape reconstruction auto-encoding. Our method encodes the field a with feature volume organized by octree applies compact multilayer perceptron network mapping features value at each position. An encoder-decoder is designed based on graph convolutions over dual nodes. The core our new convolution operator defined regular grid fused from irregular neighboring nodes different levels, which not only reduces computational memory cost nodes, but also improves performance learning. effectively details, enables fast reconstruction, exhibits good generality modeling out training categories. evaluate set tasks scenes validate its superiority other existing approaches. code, data, trained models are available https://wang-ps.github.io/dualocnn.