Optimizing Average Precision Using Weakly Supervised Data.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Computer vision tasks like action classification and object detection often require ranking samples by relevance.
• Average precision (AP) is a key performance metric, but Support Vector Machines (SVMs) typically optimize a surrogate 0-1 loss.
• While SVMs perform well in fully supervised settings, their effectiveness in weakly supervised learning, where optimizing the right metric is critical, is questionable.

Purpose of the Study:

• To investigate the hypothesis that optimizing the correct accuracy measure is crucial in weakly supervised learning.
• To develop a novel latent AP-SVM that directly optimizes an AP-based loss function in a weakly supervised context.

Main Methods:

• Proposed a novel latent AP-SVM model.
• Designed a loss function that minimizes an upper bound on the AP-based loss for weakly supervised samples.
• Evaluated the model on publicly available datasets.

Main Results:

• Demonstrated the advantage of the proposed latent AP-SVM over standard loss-based learning frameworks.
• Achieved superior performance on challenging weakly supervised problems including action classification, character recognition, and object detection.

Conclusions:

• Directly optimizing for average precision (AP) is more effective than using surrogate losses in weakly supervised computer vision.
• The proposed latent AP-SVM offers a significant improvement for practical weakly supervised learning tasks.