Figure 2

Framework of open set recognition. The left part of the blue dotted line illustrates the two training phases, the initial training phase, and the incremental training phase. The initial training phase takes the initial training set (contains the knowns only) as input, then learns and outputs an initial model, prototypes and prototype radiuses for each category. And the incremental training phase takes the incremental training set (contains both knowns and unknowns) and the outputs of the initial training phase as inputs, then detects the unknowns. Manual labeling the unknowns which are detected in the previous process. Next, the new category is dynamically incorporated in the model. Finally, fine-tuning the model with only a few samples to make the unknowns known. The right part of the dotted line illustrates two evaluation phases responding to the two training phases, the evaluation phase 1 and the evaluation phase 2. In the evaluation phase 1, the detection f-score of unknowns is measured here on the initial model trained in the initial training phase. And in the evaluation phase 2, the classification accuracy of both knowns and unknowns is measured here on the final model trained in the incremental training phase.