Visual and Semantic Knowledge Transfer for Large Scale Semi-Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Deep Convolutional Neural Network (CNN)-based object detection excels but requires extensive bounding box annotations.
• Existing methods transfer knowledge from image-level classifiers to object detectors, but can be improved.

Purpose of the Study:

• To enhance the transfer of knowledge from image classifiers to object detectors by incorporating object similarities.
• To improve semi-supervised object detection performance.

Main Methods:

• Developed a novel knowledge transfer method leveraging visual and semantic similarities between object categories.
• Applied this method to transform image-level classifiers into object detectors for categories lacking bounding box annotations.

Main Results:

• Proposed object similarity-based knowledge transfer methods significantly outperformed baseline approaches on the ILSVRC2013 detection dataset.
• Demonstrated that combining visual similarity and semantic relatedness yields complementary benefits, boosting detection performance.

Conclusions:

• Object similarity knowledge transfer is effective for semi-supervised object detection.
• The proposed methods achieve state-of-the-art results by effectively utilizing visual and semantic information to bridge the annotation gap.