Skeleton-based Action Recognition on NTU-RGB+D (Cross-subject)
93.1AccuracyPoseConv3D
Evaluation results
| Method | Links | |
|---|---|---|
| PoseConv3DModality=Joint, Sampling=1-clip2022.10 | 93.1 | |
| MotionBERTModality=Joint, Sampling=1-clip, Training=finetune2022.10 | 93 | |
| FR-HeadParam. (M)=1.3, FLOPs (G)=7.08, Power (mJ)=32.572025.12 | 92.8 | |
| PYSKLSkeleton format=3D2022.05 | 92.6 | |
| CTR-GCNSkeleton format=3D2022.05 | 92.4 | |
| CTR-GCN2023.01 | 92.4 | |
| CTR-GCNModality=Joint, Sampling=1-clip2022.10 | 92.4 | |
| CTR-GCNParam. (M)=1.46, FLOPs (G)=7.88, Power (mJ)=36.252025.12 | 92.4 | |
| Info-GCNParam. (M)=1.6, FLOPs (G)=7.36, Power (mJ)=33.862025.12 | 92.3 | |
| MS-G3DModality=Joint, Sampling=1-clip2022.10 | 91.5 | |
| MS-G3DParam. (M)=3.19, FLOPs (G)=48.88, Power (mJ)=239.512025.12 | 91.5 | |
| PA-ResGCN-B19Inference speed (sequences/(second*GPU))=38.3, Parameter number (million)=3.642020.10 | 90.9 | |
| 4s Shift-GCN2023.01 | 90.7 | |
| Shift-GCNModality=Joint, Sampling=1-clip2022.10 | 90.7 | |
| Shift-GCNFLOPs (G)=10, Power (mJ)=462025.12 | 90.7 | |
| PA-ResGCN-N51Inference speed (sequences/(second*GPU))=54.8, Parameter number (million)=1.142020.10 | 90.3 | |
| FGCN2020.03 | 90.2 | |
| ResGCN-B19 (Basic)Inference speed (sequences/(second*GPU))=44.0, Parameter number (million)=3.262020.10 | 90 | |
| MS-AAGCN2023.01 | 90 | |
| Action Capsules2023.01 | 90 | |
| DGNN2020.03 | 89.9 | |
| DGNNParameter number (million)=26.242020.10 | 89.9 | |
| MCCModality=Joint, Sampling=1-clip, Training=finetune2022.10 | 89.7 | |
| MS-G3D NetFLOPs=~8.32G, Pathways=22020.07 | 89.4 | |
| VA-fusionParameter number (million)=24.602020.10 | 89.4 | |
| NAS-GCNParameter number (million)=6.572020.10 | 89.4 | |
| AGC-LSTM2020.03 | 89.2 | |
| Dynamic GCNFLOPs=~1.99G (+~7%), Configuration=ours2020.07 | 89.2 | |
| AGC-LSTMParameter number (million)=22.892020.10 | 89.2 | |
| PL-GCNParameter number (million)=20.702020.10 | 89.2 | |
| 2s AGC-LSTM2023.01 | 89.2 | |
| MS-G3D NetFLOPs=~5.21G, Pathways=12020.07 | 89.1 | |
| ResGCN-N51 (Bottleneck)Inference speed (sequences/(second*GPU))=67.4, Parameter number (million)=0.772020.10 | 89.1 | |
| SGNsemantics=true2019.04 | 89 | |
| VA-CNNYear=20192019.04 | 88.7 | |
| 3SCNN2023.01 | 88.6 | |
| 2s-AGCN2020.03 | 88.5 | |
| 2s-AGCNYear=20192019.04 | 88.5 | |
| 2s-AGCNInference speed (sequences/(second*GPU))=22.3, Parameter number (million)=6.942020.10 | 88.5 | |
| 2s-AGCNModality=Joint, Sampling=1-clip2022.10 | 88.5 | |
| 2S-AGCNParam. (M)=3.48, FLOPs (G)=37.32, Power (mJ)=182.872025.12 | 88.5 | |
| Dynamic GCNFLOPs=~1.86G, Configuration=w/o CeN2020.07 | 88.2 | |
| MS-AAGCNFLOPs=~3.98G2020.07 | 88 | |
| SCCModality=Joint, Sampling=1-clip, Training=finetune2022.10 | 88 | |
| MotionBERTModality=Joint, Sampling=1-clip, Training=scratch2022.10 | 87.7 | |
| PB-GCN2020.03 | 87.5 | |
| AGC-LSTM (joint)Year=20192019.04 | 87.5 | |
| GR-GCNYear=20192019.04 | 87.5 | |
| PB-GCN2020.10 | 87.5 | |
| GR-GCN2020.10 | 87.5 | |
| PB-GCN2019.10 | 87.5 | |
| Signal-SGN++ (4 ensemble)Param. (M)=1.72, FLOPs (G)=6.44, SOPs (G)=1.252, Power (mJ)=1.4762025.12 | 87.2 | |
| SGNsemantics=false2019.04 | 86.9 | |
| AS-GCN2020.03 | 86.8 | |
| AS-GCNYear=20192019.04 | 86.8 | |
| AS-GCNFLOPs=~6.10G2020.07 | 86.8 | |
| AS-GCNParameter number (million)=6.992020.10 | 86.8 | |
| AS-GCN2023.01 | 86.8 | |
| UNIKModality=Joint, Sampling=1-clip, Training=finetune2022.10 | 86.8 | |
| dense-IndRNN-augAugmentation=true2019.10 | 86.7 | |
| SGNInference speed (sequences/(second*GPU))=188.0, Parameter number (million)=1.82020.10 | 86.6 | |
| HCNYear=20182019.04 | 86.5 | |
| HCN2019.10 | 86.5 | |
| HCN2023.01 | 86.5 | |
| CrosSCLRModality=Joint, Sampling=1-clip2022.10 | 86.2 | |
| Signal-SGN (4 ensemble)Param. (M)=1.74, FLOPs (G)=6.48, SOPs (G)=1.28, Power (mJ)=1.4882025.12 | 86.1 | |
| RA-GCNInference speed (sequences/(second*GPU))=18.7, Parameter number (million)=6.212020.10 | 85.9 | |
| AS-GCN+DH-TCN2020.10 | 85.3 | |
| ResNet152-3S2020.03 | 85 | |
| SR-TSL2020.03 | 84.8 | |
| SR-TSLYear=20182019.04 | 84.8 | |
| SR-TSLInference speed (sequences/(second*GPU))=14.0, Parameter number (million)=19.072020.10 | 84.8 | |
| Signal-SGN++ (Bone+Joint)Param. (M)=1.72, FLOPs (G)=3.22, SOPs (G)=0.626, Power (mJ)=0.7382025.12 | 84.5 | |
| DPRL+GCNN2020.03 | 83.5 | |
| DPRL+GCNNYear=20182019.04 | 83.5 | |
| Signal-SGN (Bone+Joint)Param. (M)=1.74, FLOPs (G)=3.24, SOPs (G)=0.628, Power (mJ)=0.7442025.12 | 82.5 | |
| Bayesian GC-LSTM2020.03 | 81.8 | |
| ST-GCN2020.03 | 81.5 | |
| ST-GCNYear=20182019.04 | 81.5 | |
| ST-GCNFLOPs=~3.56G2020.07 | 81.5 | |
| ST-GCNInference speed (sequences/(second*GPU))=42.9, Parameter number (million)=3.102020.10 | 81.5 | |
| STGCN2019.10 | 81.5 | |
| ST-GCN2023.01 | 81.5 | |
| ST-GCNModality=Joint, Sampling=1-clip2022.10 | 81.5 | |
| ST-GCNParam. (M)=3.1, FLOPs (G)=3.48, Power (mJ)=16.012025.12 | 81.5 | |
| Signal-SGN++ (Joint)Param. (M)=1.72, FLOPs (G)=1.61, SOPs (G)=0.313, Power (mJ)=0.3692025.12 | 81.3 | |
| ElAtt-GRUYear=20182019.04 | 80.7 | |
| Signal-SGN (Joint)Param. (M)=1.74, FLOPs (G)=1.62, SOPs (G)=0.314, Power (mJ)=0.3722025.12 | 80.5 | |
| Enhanced Visualization+CNN2019.10 | 80.03 | |
| Clips+CNN+MTLNYear=20172019.04 | 79.6 | |
| Clips+CNN+MTLN2019.10 | 79.57 | |
| VA-LSTMYear=20172019.04 | 79.4 | |
| VA-LSTM2023.01 | 79.4 | |
| MK-SGNParam. (M)=2.17, FLOPs (G)=7.84, SOPs (G)=0.68, Power (mJ)=0.6142025.12 | 78.5 | |
| TCN + TTN2019.10 | 77.55 | |
| Spike-driven Transformer V2Param. (M)=11.47, FLOPs (G)=38.28, SOPs (G)=2.59, Power (mJ)=2.912025.12 | 77.4 | |
| Pose conditioned STA-LSTM2019.10 | 77.1 | |
| TSA2020.10 | 76.5 | |
| JDM+CNN2019.10 | 76.2 | |
| SkeletonNet(CNN)2019.10 | 75.94 |