QBox: Partial Transfer Learning With Active Querying for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection models require extensive labeled data, which is often difficult to obtain due to privacy concerns or annotator limitations.
• Unlabeled datasets from image search engines are abundant but may contain categories beyond the target task.

Purpose of the Study:

• To propose a cost-effective partial transfer learning approach, QBox, for improving object detection models.
• To reduce labeling costs by actively querying informative bounding boxes from source data.

Main Methods:

• QBox employs a partial transfer learning strategy to query bounding box labels from source images.
• Two criteria, informativeness and transferability, are designed to measure the utility of bounding boxes for the target model.
• The approach enables annotators to label specific regions, reducing overall labeling difficulty.

Main Results:

• QBox actively queries labels for the most useful bounding boxes, requiring fewer training examples.
• Experiments on various benchmarks and a COVID-19 detection task demonstrate improved detection accuracy.
• The method achieves lower labeling costs compared to existing query strategies.

Conclusions:

• QBox offers an efficient solution for object detection in data-scarce scenarios.
• The active querying strategy effectively leverages readily available unlabeled data.
• This approach significantly reduces the burden and cost of data annotation.