Are Labels Needed for Incremental Instance Learning?

We introduce a new, self-incremental approach to visually classify object instances in this paper. Our model, \textbf{V}isual Self-\textbf{In}cremental \textbf{I}nstance \textbf{L}earning, VINIL for short, incrementally observes and classifies a single instance at a time before discarding it. The challenge in incremental instance learning is the risk of forgetting over longer learning sessions and the difficulties associated with instance labeling. Our approach addresses these challenges through three key innovations: \textit{i)}. The development of VINIL, a self-incremental learner that can learn object instances in sequence, \textit{ii)}. The use of self-supervision to eliminate the need for instance labeling, and \textit{iii)}. A thorough evaluation of VINIL against label-supervised counterparts on two large-scale benchmarks~\cite{core50,ilab20m}, showcasing its effectiveness in achieving higher accuracy while reducing forgetfulness.