Object Classification on ScanObjectNN PB-T50-RS
93.48Overall AccuracyMamba3D + Mantis
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Mamba3D + MantisBackbone=SSM, Number of trainable parameters (#TP)=0.8 (4.7%), FLOPs (G)=4.0, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 93.48 | — | |
| Mamba3DBackbone=SSM, Number of trainable parameters (#TP)=16.9 (100%), FLOPs (G)=3.9, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 92.05 | — | |
| Mamba3DPre-training Representation=Point Cloud2026.07 | 91.81 | — | |
| ReConBackbone=Attn., Number of trainable parameters (#TP)=43.6 (100%), FLOPs (G)=5.3, Evaluation Protocol=Parameter-Efficient Fine-tuning, Data Modality (Cross-modal vs Single-modal)=cross-modal, Augmentation Strategy=Simple rotational augmentation2026.05 | 90.63 | — | |
| I2P-MAEBackbone=Attn., Number of trainable parameters (#TP)=15.3, Evaluation Protocol=Full Fine-tuning, Augmentation Strategy=Simple rotational augmentation2026.05 | 90.11 | — | |
| ZigzagPointMamba + MantisBackbone=SSM, Number of trainable parameters (#TP)=0.6 (4.9%), FLOPs (G)=4.5, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 90.11 | — | |
| PointMamba + MantisBackbone=SSM, Number of trainable parameters (#TP)=0.6 (4.9%), FLOPs (G)=4.0, Evaluation Protocol=Parameter-Efficient Fine-tuning, Augmentation Strategy=Simple rotational augmentation2026.05 | 89.96 | — | |
| UtoniaParams Learn.=137.4M, Params Pct.=100%, Evaluation Protocol=fine-tuning2026.03 | 89.9 | 88.3 | |
| ReCon + GAPromptBackbone=Attn., Number of trainable parameters (#TP)=0.6 (1.4%), FLOPs (G)=5.0, Evaluation Protocol=Parameter-Efficient Fine-tuning, Data Modality (Cross-modal vs Single-modal)=cross-modal, Augmentation Strategy=Simple rotational augmentation2026.05 | 89.76 | — | |
| ReCon + DAPTBackbone=Attn., Number of trainable parameters (#TP)=1.1 (2.5%), FLOPs (G)=5.0, Evaluation Protocol=Parameter-Efficient Fine-tuning, Data Modality (Cross-modal vs Single-modal)=cross-modal, Augmentation Strategy=Simple rotational augmentation2026.05 | 89.38 | — | |
| PointMambaBackbone=SSM, Number of trainable parameters (#TP)=12.3 (100%), FLOPs (G)=3.1, Evaluation Protocol=Parameter-Efficient Fine-tuning, Augmentation Strategy=Simple rotational augmentation2026.05 | 89.31 | — | |
| PointGPT-SBackbone=Attn., Number of trainable parameters (#TP)=29.2, FLOPs (G)=5.7, Evaluation Protocol=Full Fine-tuning, Augmentation Strategy=Simple rotational augmentation2026.05 | 89.17 | — | |
| Mamba3D + PMABackbone=SSM, Number of trainable parameters (#TP)=1.3 (7.7%), FLOPs (G)=4.7, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 88.7 | — | |
| ZigzagPointMambaBackbone=SSM, Number of trainable parameters (#TP)=12.3 (100%), FLOPs (G)=3.1, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 88.65 | — | |
| ACTBackbone=Attn., Number of trainable parameters (#TP)=22.1, FLOPs (G)=4.8, Evaluation Protocol=Full Fine-tuning, Augmentation Strategy=Simple rotational augmentation2026.05 | 88.21 | — | |
| ConcertoParams Learn.=137.4M, Params Pct.=100%, Evaluation Protocol=fine-tuning2026.03 | 88.1 | 86.7 | |
| PCMBackbone=SSM, Number of trainable parameters (#TP)=34.2, Evaluation Protocol=Full Fine-tuning2026.05 | 88.1 | — | |
| SonataParams Learn.=124.8M, Params Pct.=100%, Evaluation Protocol=fine-tuning2026.03 | 87.5 | 85.4 | |
| SI-MambaBackbone=SSM, Number of trainable parameters (#TP)=12.3, FLOPs (G)=3.6, Evaluation Protocol=Full Fine-tuning2026.05 | 87.3 | — | |
| PointMamba + PMABackbone=SSM, Number of trainable parameters (#TP)=1.1 (8.9%), FLOPs (G)=5.2, Evaluation Protocol=Parameter-Efficient Fine-tuning, Augmentation Strategy=Simple rotational augmentation2026.05 | 86.71 | — | |
| Point-M2AEBackbone=Attn., Number of trainable parameters (#TP)=15.3, FLOPs (G)=3.6, Evaluation Protocol=Full Fine-tuning2026.05 | 86.43 | — | |
| Point-M2AEPre-training Representation=Point Cloud2026.07 | 86.43 | — | |
| Point-M2AEParams Learn.=12.9M, Params Pct.=100%, Evaluation Protocol=fine-tuning2026.03 | 86.4 | — | |
| Joint-MAEBackbone=Attn., Evaluation Protocol=Full Fine-tuning2026.05 | 86.07 | — | |
| Point-MAE + PointLoRABackbone=Attn., Number of trainable parameters (#TP)=0.8 (3.6%), Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 85.53 | — | |
| Point-MAEBackbone=Attn., Number of trainable parameters (#TP)=22.1 (100%), FLOPs (G)=4.8, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 85.18 | — | |
| ZigzagPointMamba + PMABackbone=SSM, Number of trainable parameters (#TP)=1.1 (8.9%), FLOPs (G)=5.7, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 85.18 | — | |
| Point-MAEPre-training Representation=Point Cloud2026.07 | 85.18 | — | |
| Point-MAE + IDPTBackbone=Attn., Number of trainable parameters (#TP)=1.7 (7.7%), FLOPs (G)=7.2, Evaluation Protocol=Parameter-Efficient Fine-tuning2026.05 | 84.94 | — | |
| Point-BERTBackbone=Attn., Number of trainable parameters (#TP)=22.1, FLOPs (G)=4.8, Evaluation Protocol=Full Fine-tuning2026.05 | 83.07 | — | |
| GaussFusionPre-training Representation=3D Gaussian2026.07 | 82.72 | — | |
| PTv3Params Learn.=124.8M, Params Pct.=100%2026.03 | 82.3 | 80 | |
| UtoniaParams Learn.=<0.2M, Params Pct.=<0.2%, Evaluation Protocol=linear probing2026.03 | 80.4 | 78.8 | |
| ConcertoParams Learn.=<0.2M, Params Pct.=<0.2%, Evaluation Protocol=linear probing2026.03 | 79.7 | 78.5 | |
| Gaussian-MAE†Pre-training Representation=3D Gaussian2026.07 | 78.87 | — | |
| Transformer + OcCoPre-training Representation=Point Cloud2026.07 | 78.79 | — | |
| PointCNNBackbone=-, Number of trainable parameters (#TP)=0.6, FLOPs (G)=0.9, Evaluation Protocol=Supervised Learning Only, Data Modality (Cross-modal vs Single-modal)=single-modal2026.05 | 78.5 | — | |
| PointCNNPre-training Representation=Point Cloud2026.07 | 78.5 | — | |
| DGCNNBackbone=-, Number of trainable parameters (#TP)=1.8, FLOPs (G)=2.4, Evaluation Protocol=Supervised Learning Only, Data Modality (Cross-modal vs Single-modal)=single-modal2026.05 | 78.1 | — | |
| DGCNNPre-training Representation=Point Cloud2026.07 | 78.1 | — | |
| PointNet++Backbone=-, Number of trainable parameters (#TP)=1.5, FLOPs (G)=1.7, Evaluation Protocol=Supervised Learning Only, Data Modality (Cross-modal vs Single-modal)=single-modal2026.05 | 77.9 | — | |
| PointNet++Pre-training Representation=Point Cloud2026.07 | 77.9 | — | |
| TransformerPre-training Representation=Point Cloud2026.07 | 77.24 | — | |
| SonataParams Learn.=<0.2M, Params Pct.=<0.2%, Evaluation Protocol=linear probing2026.03 | 76.2 | 75.2 | |
| SpiderCNNPre-training Representation=Point Cloud2026.07 | 73.7 | — | |
| PointNetBackbone=-, Number of trainable parameters (#TP)=3.5, FLOPs (G)=0.5, Evaluation Protocol=Supervised Learning Only, Data Modality (Cross-modal vs Single-modal)=cross-modal2026.05 | 68 | — | |
| PointNetPre-training Representation=Point Cloud2026.07 | 68 | — |