BooDet: Gradient Boosting Object Detection With Additive Learning-Based Prediction Aggregation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Object detection performance is challenged by the trade-off between network complexity and prediction accuracy.
• Deep neural networks have advanced object detection, but limitations persist in achieving high performance without complex architectures.

Purpose of the Study:

• To propose a novel approach, BooDet, for boosting object detection performance by aggregating predictions.
• To address the dilemma between complex networks and single-vector predictions in current object detection models.

Main Methods:

• Developed a unified module to generate multiple predictions from a single detection network.
• Formulated additive learning for aggregating predictions, reducing classification and regression losses.
• Modeled prediction optimization using gradient Boosting and weighted regression for Newton-descent directions.

Main Results:

• The BooDet approach, integrated with Cascade R-CNN, significantly improves object detection.
• Achieved a 1.3%–2.0% performance improvement over the Cascade R-CNN baseline on the COCO validation dataset.
• Attained 56.5% AP on the COCO test-dev dataset using only bounding box annotations.

Conclusions:

• BooDet effectively bootstraps classification and bounding box regression for high-performance object detection.
• The proposed prediction aggregation strategy offers a viable method for enhancing existing object detection frameworks.
• Demonstrated the effectiveness of BooDet in improving object detection accuracy and efficiency.