PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies

Benchmark Model Rank Results
3d-part-segmentation-on-shapenet-partPointNeXt#3Instance Average IoU: 87.1Class Average IoU: 85.2
3d-point-cloud-classification-on-modelnet40PointNeXt#28Overall Accuracy: 94.0Mean Accuracy: 91.1Number of params: 4.5M
3d-point-cloud-classification-on-scanobjectnnPointNeXt#35Overall Accuracy: 88.2Mean Accuracy: 86.8FLOPs: 1.64G
3d-semantic-segmentation-on-opentrench3dPointNeXt-XL#3mIoU: 70.6mAcc: 79.7Model Size: 41.5M
3d-semantic-segmentation-on-s3disPointNext#3mIoU (Area-5): 70.5mIoU (6-Fold): 74.9
semantic-segmentation-on-s3disPointNeXt-XL#12Mean IoU: 74.9mAcc: 83.0oAcc: 90.3FLOPs: 84.8G
semantic-segmentation-on-s3disPointNeXt-L#16Mean IoU: 73.9mAcc: 82.2oAcc: 89.9FLOPs: 15.2G
semantic-segmentation-on-s3dis-area5PointNeXt#24mIoU: 71.1mAcc: 77.2oAcc: 91.0Number of params: 41.6M
supervised-only-3d-point-cloud-classificationPointNeXt#6Overall Accuracy (PB_T50_RS): 87.8GFLOPs: 3.6