Figure 2.

A detailed architecture of the suggested model and CTR-GCN. We applied augmentation methods and frame sampling strategies. The pretrained weights for the CTR-GCN model were applied, and 8 basic blocks of the model were frozen during training. The Grad-CAM was generated from the gradients and feature map from the last block. The Grad-CAM was then interpolated to align with the input frames. A basic block of CTR-GCN consists of 3 CTR-GCs, which use temporal pooling to aggregate temporal features of skeleton graph sequences and pairwise subtraction and concatenation for correlation modeling between skeletal joints. C: channel dimension of the data; CTR-GCN: channel-wise topology refinement graph convolution networks; GMS: gross motor skills; Grad-CAM: gradient-weighted class activation mapping; MLP: multilayer perceptron; N: number of skeletal joints; T: temporal dimension of the data; Tanh: hyperbolic tangent function; XGBoost: extreme gradient boosting.