Learning optimal embedded cascades.

Area of Science:

• Computer Vision
• Machine Learning
• Embedded Systems

Background:

• Designing optimal embedded object detector cascades presents challenges in configuration and stage optimization.
• Achieving a balance between classification accuracy and processing speed under detection rate constraints is crucial for real-time applications.

Purpose of the Study:

• To develop automated methods for optimizing embedded object detector cascades.
• To address the dual challenges of cascade configuration and individual stage performance.
• To ensure a target detection rate is met throughout the optimization process.

Main Methods:

• Introduced RCBoost, a constrained optimization approach using a barrier penalty method to optimize cascade stages while meeting detection rate constraints.
• Developed ECBoost to search for optimal cascade configurations, balancing classification risk and speed.
• Combined RCBoost and ECBoost into RCECBoost for fully automated optimization of both configuration and stages.

Main Results:

• RCBoost enables the design of embedded cascades with known configurations without extensive cross-validation.
• RCECBoost successfully optimizes cascade configuration and stages under a detection rate constraint.
• Experiments demonstrated superior accuracy-speed tradeoffs for face, car, pedestrian, and panda detection compared to prior methods.

Conclusions:

• The proposed RCECBoost algorithm offers a fully automated solution for designing high-performance embedded object detector cascades.
• This approach significantly improves the accuracy-speed tradeoff, making it suitable for resource-constrained environments.
• The methods are validated across diverse object detection tasks, highlighting their general applicability.