Learning Near-Optimal Cost-Sensitive Decision Policy for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Popular object detectors like AdaBoost, SVM, and deformable part-based models (DPM) require efficient computation for real-time applications.
• Computational efficiency is crucial alongside accuracy in object detection tasks.

Purpose of the Study:

• To develop a cost-sensitive decision policy for object detection that minimizes computational cost and detection errors.
• To improve the efficiency of object detection algorithms without sacrificing accuracy.

Main Methods:

• Formulated an empirical risk function balancing computation cost and detection loss (false alarms, missed detections).
• Developed a decision policy using a sequence of two-sided thresholds for early rejection or acceptance.
• Optimized the upper bound of the risk function using dynamic programming, achieving a near-optimal policy.

Main Results:

• The proposed decision policy significantly enhances computational efficiency compared to state-of-the-art cascade methods.
• Empirical validation demonstrated a tight upper bound for the risk function, confirming the policy's near-optimality.
• Achieved similar detection accuracy while substantially improving efficiency across popular detection benchmarks.

Conclusions:

• The novel decision policy offers a near-optimal approach to balancing computational cost and detection performance.
• This method provides a significant advancement in efficient object detection for real-time systems.
• The dynamic programming optimization provides a practical solution for complex risk functions in object detection.