Improving Open-world Continual Learning under the Constraints of Scarce Labeled Data