VoxelFix: Post-Hoc Semantic Correction of Completed 3D Voxel Maps

To address perception and mapping errors in the completed 3D voxel map, the research team proposed a post-processing semantic correction method called VoxelFix. Based on graph models, this method directly corrects voxel labels using local geometry and neighboring semantic information while maintaining the geometric structure and occupancy information. Experiments were conducted by creating a dataset with category confusion in the OccuFly map and evaluating it on maps generated by four independently trained 2D segmentation models. The results showed that VoxelFix increased the average intersection-over-union score (mIoU) by 4.23 to 5.00 percentage points, and significant improvements were observed in categories such as trees, roofs, and walls. Additionally, this method demonstrated good generalization ability in out-of-distribution scenarios reconstructed independently.