Deep Ensemble Learning for Human Action Recognition in Still Images

Figure 5

The whole framework for DELWO model. (1) For training phase, training data are fed into multiple deep models, then the GAP layers are concatenated into one long layer; finally, the fused information derived from multiple deep models is fed into the classifier module and conducts weight optimization. (2) For testing phase, testing data are fed into the training model to get the final prediction probability.