Specific Action Recognition Method based on Unbalanced Dataset
Yuan Liu, Xiangdong You · 2019
Specific Action Recognition is a mission that refers to identifying a certain type of human from a video, e.g., fighting, chasing. In the actual application, the number of negative samples is far more than positive samples. This paper focus on the specific action recognition of the human body based on extremely unbalanced dataset, and proposes A specific action recognition method that can be applied to unbalanced dataset. The input of the method are video frames. After extracting image features by 2D convolutional networks and stacking the image feature maps, the time domain features are fused by a small number of 3D convolutional networks. In order to solve the problem that dataset is imbalanced, negative videos are under-sampled firstly, then focal loss and weighted cross entropy loss are used to guide the model training. This model is trained on Kinetics-600 dataset and self-collecting specific action dataset. Finally, average accuracy on the self-collected dataset is 95.47%.