Few-Example Object Detection with Model Communication.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection typically requires large labeled datasets.
• Few-example object detection addresses the challenge of limited labeled data.
• Generating high-quality training samples from unlabeled data is crucial.

Purpose of the Study:

• To develop a method for few-example object detection.
• To efficiently generate trustworthy training samples from a large pool of unlabeled images.
• To achieve competitive performance with minimal labeled data.

Main Methods:

• An iterative approach combining model training and high-confidence sample selection.
• Progressive selection of training samples, starting with easy ones and progressing to challenging ones.
• Ensemble of multiple detection models to enhance precision and recall of generated samples.

Main Results:

• The proposed method significantly outperforms single-model baselines and traditional ensemble methods.
• Achieves competitive results compared to state-of-the-art weakly-supervised methods.
• Demonstrates effectiveness on benchmark datasets like PASCAL VOC'07, MS COCO'14, and ILSVRC'13.

Conclusions:

• Few-example object detection is feasible and effective with the proposed iterative sampling and ensemble strategy.
• The method offers a powerful alternative when labeled data is scarce.
• This approach significantly reduces the need for extensive manual data labeling in object detection tasks.