Self-improving object detection via disagreement reconciliation

TL;DR

This paper introduces a self-supervised method for adapting object detectors to new environments by reconciling disagreements among pseudo-labels, leading to improved detection performance without human annotations.

Contribution

It proposes a novel disagreement reconciliation mechanism for self-supervised fine-tuning of object detectors in unseen environments.

Findings

Improves mAP by 2.66% over baseline

Outperforms state-of-the-art methods without ground-truth labels

Enables autonomous adaptation to new environments

Abstract

Object detectors often experience a drop in performance when new environmental conditions are insufficiently represented in the training data. This paper studies how to automatically fine-tune a pre-existing object detector while exploring and acquiring images in a new environment without relying on human intervention, i.e., in a self-supervised fashion. In our setting, an agent initially explores the environment using a pre-trained off-the-shelf detector to locate objects and associate pseudo-labels. By assuming that pseudo-labels for the same object must be consistent across different views, we devise a novel mechanism for producing refined predictions from the consensus among observations. Our approach improves the off-the-shelf object detector by 2.66% in terms of mAP and outperforms the current state of the art without relying on ground-truth annotations.