Abstract
In this paper, we learn to classify visual object instances, incrementally and via self-supervision (self-incremental). Our learner observes a single instance at a time, which is then discarded from the dataset. Incremental instance learning is challenging, since longer learning sessions exacerbate forgetfulness, and labeling instances is cumbersome. We overcome these challenges via three contributions: \textit{i).} We propose VINIL, a self-incremental learner that can learn object instances sequentially, \textit{ii).} We equip VINIL with self-supervision to by-pass the need for instance labelling, \textit{iii).} We compare VINIL to label-supervised variants on two large-scale benchmarks~\cite{core50,ilab20m}, and show that VINIL significantly improves accuracy while reducing forgetfulness.
Abstract (translated)
在本文中,我们通过学习逐步地对视觉对象实例进行分类(自增量学习),并使用自我监督(自我增量学习)方法进行分类。我们的学习器一次性地观察一个实例,并将其从数据集中删除。自增量学习是挑战性的,因为更长的学习周期会加重遗忘,标注实例很不方便。我们通过三项贡献克服了这些挑战: extit{i).} 我们提出了VINIL,它是一种自增量学习器,可以逐步地对对象实例进行分类。 extit{ii).} 我们为VINIL配备了自我监督,以绕过实例标注的需求。 extit{iii).} 我们比较了VINIL与标注监督的变体,在两个大型基准问题上~$cite{core50,ilab20m}$,并表明VINIL significantly improves accuracy,同时减少了遗忘。
URL
https://arxiv.org/abs/2301.11417
