Are Labels Needed for Incremental Instance Learning?

TL;DR

This paper introduces VINIL, a self-supervised incremental learning method for visual object instances that learns sequentially without labels, addressing forgetfulness and reducing the need for manual annotations.

Contribution

We propose VINIL, a novel self-incremental, self-supervised learner for object instance classification that outperforms label-supervised methods on large-scale benchmarks.

Findings

VINIL significantly improves accuracy over label-supervised methods.

VINIL reduces forgetfulness in incremental learning sessions.

Self-supervision enables learning without manual labels.

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: i. We propose VINIL, a self-incremental learner that can learn object instances sequentially, ii. We equip VINIL with self-supervision to by-pass the need for instance labelling, iii. We compare VINIL to label-supervised variants on two large-scale benchmarks, and show that VINIL significantly improves accuracy while reducing forgetfulness.