Fig. 1. Network architecture: four local feature encoding blocks (dotted box above, with detail structure of a single local feature encoding block shown in lower left dotted box) are assembled recursively to encode local features, then the four outputs (lower right solid box) are concatenated to be input to decoding blocks

Fig. 2. Example of edge convolution operation

Fig. 3. S3DIS data set

Fig. 4. Accuracy (dotted line) and loss function (solid line) curves

Fig. 5. Segmentation accuracy of S3DIS data set for different k

Fig. 6. Sample semantic segmentation results of S3DIS data set. (a) Ground truth; (b) segmentation result of PointNet++; (c) segmentation result of DGPoint

Table 1. Comparison of semantic segmentation accuracy on entire S3DIS data set

Table 2. IoU of some semantic class in S3DIS data set