Few-shot object classification on ModelNet40 (test)
96.4Accuracy (5-way, 10-shot)Mamba3D
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Mamba3DPre-training Representation=Point Cloud2026.07 | 96.4 | 98.2 | 92.4 | 95.2 | |
| Point-MAEPre-training/Initialization strategy=MAE-style pre-training2022.03 | 96.3 | 97.8 | 92.6 | 95 | |
| Point-MAEPre-training Representation=Point Cloud2026.07 | 96.3 | 97.8 | 92.6 | 95 | |
| GaussFusionPre-training Representation=3D Gaussian2026.07 | 96.1 | 96.5 | 88.9 | 93.1 | |
| MaskPointPre-training=Masked Discrimination2022.03 | 95 | 97.2 | 91.4 | 93.4 | |
| Point-BERTPre-training/Initialization strategy=BERT-style pre-training2022.03 | 94.6 | 96.3 | 91 | 92.7 | |
| Point-BERTPre-training=Point-BERT2022.03 | 94.6 | 96.3 | 91 | 92.7 | |
| Point-BERTPre-training Representation=Point Cloud2026.07 | 94.6 | 96.3 | 91 | 92.7 | |
| Transformer-OcCoPre-training/Initialization strategy=OcCo pre-training2022.03 | 94 | 95.9 | 89.4 | 92.4 | |
| PointViT-OcCoBackbone=ViT, Pre-training=OcCo2022.03 | 94 | 95.9 | 89.4 | 92.4 | |
| Transformer+OcCoPre-training Representation=Point Cloud2026.07 | 94 | 95.9 | 89.4 | 92.4 | |
| Gaussian-MAEPre-training Representation=3D Gaussian, reproducibility=reproduced results under our implementation2026.07 | 94 | 95.1 | 87.3 | 92.2 | |
| DGCNN + CrossPointBackbone=DGCNN2022.03 | 92.5 | 94.9 | 83.6 | 87.9 | |
| DGCNN-OcCoBackbone=DGCNN, Pre-training=OcCo2022.03 | 91.9 | 93.9 | 86.4 | 91.3 | |
| DGCNNBackbone=DGCNN2022.03 | 91.8 | 93.4 | 86.3 | 90.9 | |
| PointNet + CrossPointBackbone=PointNet2022.03 | 90.9 | 93.5 | 84.6 | 90.2 | |
| DGCNN + OcCoBackbone=DGCNN2022.03 | 90.6 | 92.5 | 82.9 | 86.5 | |
| DGCNN-OcCoPre-training/Initialization strategy=OcCo pre-training2022.03 | 90.6 | 92.5 | 82.9 | 86.5 | |
| DGCNN-rand+OcCoPre-training Representation=Point Cloud2026.07 | 90.6 | 92.5 | 82.9 | 86.5 | |
| PointNet + OcCoBackbone=PointNet2022.03 | 89.7 | 92.4 | 83.9 | 89.7 | |
| Transformer-randPre-training/Initialization strategy=Random initialization2022.03 | 87.8 | 93.3 | 84.6 | 89.4 | |
| PointViTBackbone=ViT, Pre-training=None (Scratch)2022.03 | 87.8 | 93.3 | 84.6 | 89.4 | |
| TransformerPre-training Representation=Point Cloud2026.07 | 87.8 | 93.3 | 84.6 | 89.4 | |
| PointNet + JigsawBackbone=PointNet2022.03 | 66.5 | 69.2 | 56.9 | 66.5 | |
| PointCNN2022.03 | 65.4 | 68.6 | 46.6 | 50 | |
| RSCNN2022.03 | 65.4 | 68.6 | 46.6 | 50 | |
| PointNet + cTreeBackbone=PointNet2022.03 | 63.2 | 68.9 | 49.2 | 50.1 | |
| DGCNN + cTreeBackbone=DGCNN2022.03 | 60 | 65.7 | 48.5 | 53 | |
| 3D-GAN2022.03 | 55.8 | 65.8 | 40.3 | 48.4 | |
| PointNet + RandBackbone=PointNet2022.03 | 52 | 57.8 | 46.6 | 35.2 | |
| 3D-PointCapsNet2022.03 | 42.3 | 53 | 38 | 27.2 | |
| Latent-GAN2022.03 | 41.6 | 46.2 | 32.9 | 25.5 | |
| PointNet++2022.03 | 38.5 | 42.4 | 23.1 | 18.8 | |
| DGCNN + JigsawBackbone=DGCNN2022.03 | 34.3 | 42.2 | 26 | 29.9 | |
| FoldingNet2022.03 | 33.4 | 35.8 | 18.6 | 15.4 | |
| DGCNN + RandBackbone=DGCNN2022.03 | 31.6 | 40.8 | 19.9 | 16.9 | |
| DGCNN-randPre-training/Initialization strategy=Random initialization2022.03 | 31.6 | 40.8 | 19.9 | 16.9 | |
| DGCNN-randPre-training Representation=Point Cloud2026.07 | 31.6 | 40.8 | 19.9 | 16.9 |