Introduction

Problem Background

Object detection is a core computer vision task that requires a model to both recognize objects and localize them with bounding boxes. It is widely used in applications such as autonomous driving, remote sensing, surveillance, robotics, and industrial inspection.

However, training strong object detectors usually depends on large-scale bounding-box annotations. Unlike image classification, where each image may only need one label, object detection requires annotators to draw precise boxes around every object instance.

Annotation Bottleneck

Bounding-box annotation is expensive and time-consuming. The cost grows quickly when the dataset is large, when images contain many objects, or when the target domain requires specialized knowledge.

Active learning helps reduce this cost by selecting only a subset of informative images for annotation. However, even after active learning selects fewer images, humans still need to draw bounding boxes from scratch on the selected samples.

Limitations of Current Active Learning Pipelines

Existing active learning methods for object detection mainly focus on selecting which images should be labeled. Many of them rely on detector confidence scores or diversity-based criteria, but these signals do not fully capture the errors that matter most for detection.

In practice, detector confidence scores may be poorly calibrated, negative images are often underused, and high-confidence false positives may be missed because the model appears confident about them.

Project Goal

This project aims to make active learning for object detection more annotation-efficient. The key question is how to reduce both the number of queried images and the effort required to annotate each selected image.

To address this problem, we explore how multiple foundation models can assist different stages of the active learning pipeline, while keeping humans in the loop for lightweight binary verification instead of full manual box drawing.