Action Assessment by Joint Relation Graphs

VI-Net—View-Invariant Quality of Human Movement Assessment

“…All networks were trained for 20 epochs using stochastic gradient descent optimization with initial learning rate of 0.001, and batch size 5. To evaluate the performance of the proposed method, we used Spearman’s rank correlation as used in References [ 1 , 3 , 4 ].…”

“…Although their method estimated action scores better than human non-experts, it was less accurate than human expert judgments. More recently, deep learning methods have been deployed to assess the quality of sport actions in RGB-only data, such as References [ 1 , 2 , 3 , 4 , 32 , 33 ]. For example, Li et al [ 1 ] divided a video into several clips to extract their spatio-temporal features by differently weighted C3D [ 34 ] networks and then concatenated the features for input to another C3D network to predict action scores.…”

“…The authors of References [ 4 , 33 ] propose I3D [ 35 ] based methods to analyse human movement. Pan et al [ 4 ] combine I3D features with pose information by building join relation graphs to predict score movement. Tang et al [ 33 ] proposed a novel loss function which addresses the intrinsic score distribution uncertainty of sport actions in the decisions by different judges.…”

“…Beyond the realms of action detection and recognition, action analysis includes the automatic assessment of the quality of human action or movement, for example, in sports action analysis [ 1 , 2 , 3 , 4 ], skill assessment [ 5 , 6 ], and patient rehabilitation movement analysis [ 7 , 8 ]. For example, in the latter application, clinicians observe patients performing specific actions in the clinic, such as walking or sitting-to-standing, to establish an objective marker for their level of functional mobility.…”

“…Recent RGB (red, green, blue) based action analysis methods, such as References [ 2 , 3 , 4 , 6 ], are not able to deal with view-invariance when applied to viewpoints significantly different to their training data. To achieve some degree of invariance, some works such as References [ 7 , 8 , 9 , 10 , 11 , 12 , 13 ], have made use of 3D human pose obtained from (i) Kinect, (ii) motion capture, or (iii) 3D pose estimation methods.…”