Action Recognition on NTU RGB+D 60 (X-sub)
97.9AccuracyMSG3D + DG-STGCN + ProtoGCN
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| MSG3D + DG-STGCN + ProtoGCNMethod=Ensemble2026.02 | 97.9 | — | — | |
| RGBPose-Conv3DModality=RGB + Pose, Fusion Strategy=early+late fusion, Backbone=ResNet-502021.04 | 97 | — | — | |
| PoseC3DPre-training=true, Feature: RGB=true, Feature: Est. Pose=true, Feature: Annot. Pose=false2022.10 | 97 | — | — | |
| PoseC3DModality=RGB + Estimated Pose, Pre-trained=true2022.12 | 97 | — | — | |
| PoseConv3DModality=B+RGB, GFLOPs=41.8*, Params(M)=31.6*2026.01 | 97 | — | — | |
| BHaRNet-MModality=J+B+JM+BM+RGB, GFLOPs=13.0, Params(M)=22.22026.01 | 97 | — | — | |
| SkeleTGFLOPs=9.6, Params(M)=5.22026.01 | 97 | — | — | |
| BHaRNet-MModality=J+B+RGB, GFLOPs=7.6, Params(M)=16.72026.01 | 96.8 | — | — | |
| Ours (BHaRNet-P)GFLOPs=6.8, Params(M)=9.72026.01 | 96.8 | — | — | |
| Ours (BHaRNet-E)GFLOPs=10.9, Params(M)=112026.01 | 96.7 | — | — | |
| VPN++ (3D pose)Modality=RGB + Pose, Pre-trained=true2022.12 | 96.6 | — | — | |
| MMNetModality=J+B+RGB, GFLOPs=89.2, Params(M)=34.22026.01 | 96.6 | — | — | |
| BHaRNet-E‡GFLOPs=3.3, Params(M)=2.82026.01 | 96.6 | — | — | |
| BHaRNet-P‡GFLOPs=3.4, Params(M)=4.92026.01 | 96.6 | — | — | |
| Ours (BHaRNet-B)GFLOPs=6.6, Params(M)=5.52026.01 | 96.6 | — | — | |
| BHaRNet-B‡GFLOPs=9.6, Params(M)=5.22026.01 | 96.5 | — | — | |
| BHaRNet-M†Modality=J+B+RGB, GFLOPs=7.6, Params(M)=16.72026.01 | 96.3 | — | — | |
| BHaRNet-P†GFLOPs=3.4, Params(M)=4.92026.01 | 96.3 | — | — | |
| BHaRNet-E†GFLOPs=5.4, Params(M)=5.52026.01 | 96.2 | — | — | |
| EPAM-NetModality=J+RGB, GFLOPs=8.1, Params(M)=2.52026.01 | 96.1 | — | — | |
| BHaRNet-B†GFLOPs=3.3, Params(M)=2.82026.01 | 96.1 | — | — | |
| MMNetModality=RGB + Pose, Pre-trained=true2022.12 | 96 | — | — | |
| Method [14]Modality=RGB + Pose2021.04 | 95.7 | — | — | |
| REASON+LLMCategory=LLM, Mod.=J, Two-stage=N, Interpret.=Y2026.05 | 95.6 | — | — | |
| VPNPre-training=false, Feature: RGB=true, Feature: Est. Pose=false, Feature: Annot. Pose=true2022.10 | 95.5 | — | — | |
| VPNModality=RGB + Pose, Pre-trained=true2022.12 | 95.5 | — | — | |
| X3D-UGTGFLOPs=9.15, Params (M)=0.96, Modality=RGB-only2026.02 | 95.21 | — | — | |
| SUGARCategory=LLM, Mod.=J, Two-stage=Y, Interpret.=Y2026.05 | 95.2 | — | — | |
| LLM-ARCategory=LLM, Mod.=J, Two-stage=Y, Interpret.=Y2026.05 | 95 | — | — | |
| EPAM-NetGFLOPs=4.61, Params (M)=2.06, Modality=RGB-only2026.02 | 94.8 | — | — | |
| 3MformerGFLOPs=58.5, Params(M)=6.72026.01 | 94.8 | — | — | |
| EPP-NetModality=Skeleton+Parsing2024.01 | 94.7 | — | — | |
| EPP-Net (J+B+P)Modality=Skeleton+Parsing2024.01 | 94.6 | — | — | |
| TSGCNeXt (new4s)Streams=4s2023.04 | 94.47 | — | — | |
| OursModality=RGB + Pose, Pre-trained=false2022.12 | 94.3 | — | — | |
| REASONCategory=GCN, Mod.=J+B+JM+BM, Two-stage=N, Interpret.=Y2026.05 | 94.3 | — | — | |
| TSGCNeXt (2s)Streams=2s2023.04 | 94.27 | — | — | |
| PoseConv3D (J+L)Input Modality=joint/limb, Skeleton Source=high-quality 2D skeletons (Sec 3.1)2021.04 | 94.1 | — | — | |
| VT-LVLM-ARModality=RGB-only2026.02 | 94.1 | — | — | |
| PoseConv3DGFLOPs=31.8, Params(M)=42026.01 | 94.1 | — | — | |
| pi-ViTModality=J+RGB, GFLOPs=590.0, Params(M)=121.42026.01 | 94 | — | — | |
| ProtoGCNEnsemble size=62024.11 | 93.8 | — | — | |
| HiODBackbone=GCN, Evaluation Protocol=FT.2026.02 | 93.8 | — | — | |
| ProtoGCNGFLOPs=43.4, Params(M)=24.92026.01 | 93.8 | — | — | |
| PoseConv3D (J)Input Modality=joint, Skeleton Source=high-quality 2D skeletons (Sec 3.1)2021.04 | 93.7 | — | — | |
| PoseC3DModality=Estimated Pose, Pre-trained=true2022.12 | 93.7 | — | — | |
| PYSKLBackbone=CNN, Evaluation Protocol=FT.2026.02 | 93.7 | — | — | |
| Hyper-GCNCategory=GCN, Mod.=J+B+JM+BM, Two-stage=N, Interpret.=N2026.05 | 93.7 | — | — | |
| TSGCNeXt (1s)Streams=1s2023.04 | 93.65 | — | — | |
| ACLNetPublication=This Paper2026.01 | 93.6 | — | — | |
| DeGCNGFLOPs=6.9, Params(M)=5.62026.01 | 93.6 | — | — | |
| VPNModality=Skeleton+RGB, Source=ECCV'202024.01 | 93.5 | — | — | |
| LA-GCNModality=Skeleton+Text, Source=arXiv'232024.01 | 93.5 | — | — | |
| VPNTraining Modality=Multi-modality, Inference Modality=Multi-modality, Multi-stream ensemble=true2024.07 | 93.5 | — | — | |
| MMCLTraining Modality=Multi-modality, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 93.5 | — | — | |
| ProtoGCNEnsemble size=42024.11 | 93.5 | — | — | |
| SkateFormerCategory=Trans., Mod.=J+B+JM+BM, Two-stage=N, Interpret.=N2026.05 | 93.5 | — | — | |
| HD-GCN2024.11 | 93.4 | — | — | |
| JT-GraphFormer2024.11 | 93.4 | — | — | |
| DVANetGFLOPs=25.5, Params (M)=40.5, Modality=RGB-only2026.02 | 93.4 | — | — | |
| HD-GCNPublication=ICCV 20232026.01 | 93.4 | — | — | |
| ST-GCN++ (4s)Conference=CVPR2022, Streams=4s2023.04 | 93.2 | — | — | |
| Zhu et al. (2019)GFLOPs=51.7, Params (M)=33.0, Modality=RGB-only2026.02 | 93.2 | — | — | |
| TSGCNeXt (new4s)Streams=4s2023.04 | 93.16 | — | — | |
| MAMPBackbone=Transformer, evaluation protocol=fine-tuned2023.08 | 93.1 | — | — | |
| SkeletonGCLTraining Modality=Pose, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 93.1 | — | — | |
| DS-GCN2024.11 | 93.1 | — | — | |
| BlockGCN2024.11 | 93.1 | — | — | |
| BlockGCNBackbone=GCN, Evaluation Protocol=FT.2026.02 | 93.1 | — | — | |
| DS-GCNPublication=AAAI 20242026.01 | 93.1 | — | — | |
| BlockGCNPublication=CVPR 20242026.01 | 93.1 | — | — | |
| VA-ARPublication=AAAI 20252026.01 | 93.1 | — | — | |
| BlockGCNGFLOPs=6.5, Params(M)=5.22026.01 | 93.1 | — | — | |
| DS-GCNCategory=GCN, Mod.=J+B+JM+BM, Two-stage=N, Interpret.=N2026.05 | 93.1 | — | — | |
| BlockGCNCategory=GCN, Mod.=J+B+JM+BM, Two-stage=N, Interpret.=N2026.05 | 93.1 | — | — | |
| InfoGCNPre-training=false, Feature: RGB=false, Feature: Est. Pose=false, Feature: Annot. Pose=true2022.10 | 93 | — | — | |
| InfoGCNModality=Pose, Pre-trained=false2022.12 | 93 | — | — | |
| InfoGCNModality=Skeleton, Source=CVPR 222024.01 | 93 | — | — | |
| SAP-CTRTraining Modality=Pose, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 93 | — | — | |
| InfoGCNTraining Modality=Pose, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 93 | — | — | |
| InfoGCN2024.11 | 93 | — | — | |
| ProtoGCNEnsemble size=22024.11 | 93 | — | — | |
| InfoGCNPublication=CVPR 20222026.01 | 93 | — | — | |
| InfoGCNGFLOPs=10.0*, Params(M)=9.42026.01 | 93 | — | — | |
| HD-GCN*Category=GCN, Mod.=J+B+J'+B', Two-stage=N, Interpret.=N2026.05 | 93 | — | — | |
| TSGCNeXt (2s)Streams=2s2023.04 | 92.91 | — | — | |
| PSUMNetParams. (M)=2.8, FLOPs (G)=2.72022.08 | 92.9 | — | — | |
| HyperformerType=Transformer, Parameters(M)=2.6M, Modality=4 streams2022.11 | 92.9 | — | — | |
| PSUMNetModality=Skeleton, Source=ECCV'222024.01 | 92.9 | — | — | |
| LSTModality=Skeleton+Text, Source=ICCV'232024.01 | 92.9 | — | — | |
| PSUMNetTraining Modality=Pose, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 92.9 | — | — | |
| KoopmanTraining Modality=Pose, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 92.9 | — | — | |
| LSTTraining Modality=Multi-modality, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 92.9 | — | — | |
| GAP2024.11 | 92.9 | — | — | |
| GAPCategory=LLM, Mod.=J+B+JM+BM, Two-stage=N, Interpret.=N2026.05 | 92.9 | — | — | |
| ST-GCN++ (2s)Conference=CVPR2022, Streams=2s2023.04 | 92.8 | — | — | |
| TCA-GCNfusion=four-stream2022.05 | 92.8 | — | — | |
| FR-HeadTraining Modality=Pose, Inference Modality=Pose, Multi-stream ensemble=true2024.07 | 92.8 | — | — | |
| SkeletonGCL2024.11 | 92.8 | — | — | |
| FR-Head2024.11 | 92.8 | — | — |