大模型辅助的噪声鲁棒无监督骨架表示学习

蔡蕊,郭钦涵,张子豪,董建锋,张荣,刘宝龙,王勋

Large model-assisted noise-robust unsupervised skeleton representation learning

Rui CAI,Qinhan GUO,Zihao ZHANG,Jianfeng DONG,Rong ZHANG,Baolong LIU,Xun WANG

表 1 NTU RGB+D 60、NTU RGB+D 120以及PKU-MMD II数据集上所提方法与其他方法的骨架动作识别性能比较

Tab.1 Comparison of skeleton-based action recognition performance of proposed method and other methods on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD II datasets

方法 A/% N60 x-sub N60 x-view N120 x-sub N120 x-setup P-II x-sub AimCLR++[21] 77.2 81.5 65.5 67.8 41.2 ViA[22] 78.1 85.8 69.2 66.9 — 3s-C-F[23] 79.1 87.2 69.2 70.8 49.8 HaLP[24] 79.7 86.8 71.1 72.2 43.5 HiCo[2] 81.1 88.6 72.8 74.1 49.4 HSTM-Transformer[17] 81.8 90.4 73.5 74.6 46.9 UmURL[25] 82.3 89.8 73.5 74.3 52.1 Skeleton-logoCLR[26] 82.4 87.2 72.8 73.5 54.7 KTCL[3] 82.4 89.4 74.4 74.5 55.5 RMMD[4] 83.0 90.5 75.2 75.8 50.6 PCM3[27] 83.9 90.4 76.5 77.5 51.5 Lin等[28] 83.9 90.3 75.7 77.2 — C2VL[14] 84.4 89.8 76.0 78.7 52.6 本研究方法 85.8 92.5 78.1 79.6 54.2